waterfall 0.1.0__py3-none-any.whl

waterfall/WatermarkerBase.py

@@ -0,0 +1,316 @@
+import gc
+import logging
+import os
+import time
+from collections import defaultdict
+from functools import partial
+from multiprocessing import Pool
+from typing import List, Tuple, Optional
+
+import numpy as np
+import torch
+from scipy.sparse import csr_matrix, vstack
+from tqdm import tqdm
+from transformers.modeling_utils import PreTrainedModel
+from transformers.tokenization_utils_base import PreTrainedTokenizerBase
+from transformers.generation.logits_process import LogitsProcessor, TopKLogitsWarper, TopPLogitsWarper
+
+from waterfall.permute import Permute
+from waterfall.WatermarkingFn import WatermarkingFn
+from waterfall.WatermarkingFnFourier import WatermarkingFnFourier
+
+class PerturbationProcessor(LogitsProcessor):
+    def __init__(self,
+                 N : int = 32000,     # Vocab size
+                 id : int = 0,        # Watermark ID
+                 ) -> None:
+
+        self.id = id
+        self.N = N
+        self.init_token_count = None
+        self.phi = np.ones(N)
+        self.n_gram = 2
+
+        self.skip_watermark = False
+
+        self.permute = Permute(self.N)
+
+    def reset(self, n_gram : int = 2) -> None:
+        self.n_gram = n_gram
+        self.init_token_count = None
+        if np.allclose(self.phi,np.median(self.phi)):
+            self.skip_watermark = True
+            logging.warning(f"Generating without watermark as watermarking function is flat")
+        else:
+            self.skip_watermark = False
+
+    def set_phi(self, phi : np.ndarray) -> None:
+        self.phi = phi
+
+    def __call__(self, input_ids: torch.LongTensor,
+                 scores: torch.FloatTensor) -> torch.FloatTensor:
+
+        if self.skip_watermark:
+            return scores
+
+        if self.init_token_count is None:
+            self.init_token_count = input_ids.shape[1]
+
+        # Insufficient tokens generated for n-gram
+        if self.init_token_count + self.n_gram - 1 > input_ids.shape[1]:
+            return scores
+
+        prev_tokens = input_ids[:,-self.n_gram+1:].cpu().numpy()
+        permutations = [self.permute.get_permutation(prev_tokens[i,:], self.id, cache=True) for i in range(prev_tokens.shape[0])]
+
+        scores[:,:self.N] += torch.tensor(self.phi[permutations],
+                                          device=scores.device,
+                                          dtype=scores.dtype)
+        return scores
+
+def indices_to_counts(N : int, dtype : np.dtype, indices : np.ndarray) -> csr_matrix:
+    counts = csr_matrix([np.bincount(j, minlength=N).astype(dtype) for j in indices])
+    return counts
+
+class Watermarker:
+    def __init__(self,
+                 tokenizer : PreTrainedTokenizerBase,
+                 model : Optional[PreTrainedModel] = None,
+                 id : int = 0,
+                 kappa : float = 6,
+                 k_p : int = 1,
+                 n_gram : int = 2,
+                 watermarkingFnClass = WatermarkingFnFourier
+                 ) -> None:
+        assert kappa >= 0, f"kappa must be >= 0, value provided is {kappa}"
+
+        assert (model is None) or isinstance(model, PreTrainedModel), f"model must be a transformers model, value provided is {type(model)}" # argument order for tokenizer and model were swapped since the original code
+
+        self.tokenizer = tokenizer
+        self.model = model
+        self.id = id
+        self.k_p = k_p
+        self.n_gram = n_gram
+        self.kappa = kappa
+
+        self.N = self.tokenizer.vocab_size
+        self.logits_processor = PerturbationProcessor(N = self.N, id = id)
+
+        self.compute_phi(watermarkingFnClass)
+
+    def compute_phi(self, watermarkingFnClass = WatermarkingFnFourier) -> None:
+        self.watermarking_fn: WatermarkingFn = watermarkingFnClass(id = id, k_p = self.k_p, N = self.N, kappa = self.kappa)
+        self.phi = self.watermarking_fn.phi
+
+        self.logits_processor.set_phi(self.phi)
+
+    def generate(
+            self,
+            prompt : Optional[str] = None,
+            tokd_input : Optional[torch.Tensor] = None,
+            n_gram : Optional[int] = None,
+            max_new_tokens : int = 1000,
+            return_text : bool =True,
+            return_tokens : bool =False,
+            return_scores : bool =False,
+            do_sample : bool =True,
+            **kwargs
+            ) -> List[str] | dict:
+
+        assert self.model is not None, "Model is not loaded. Please load the model before generating text."
+
+        if n_gram is None:
+            n_gram = self.n_gram
+        if tokd_input is None:
+            assert prompt is not None, "Either prompt or tokd_input must be provided."
+            tokd_input = self.tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
+        tokd_input = tokd_input.to(self.model.device)
+        logits_processor = []
+        if "top_k" in kwargs and kwargs["top_k"] is not None and kwargs["top_k"] != 0:
+            logits_processor.append(TopKLogitsWarper(kwargs.pop("top_k")))
+        if "top_p" in kwargs and kwargs["top_p"] is not None and kwargs["top_p"] < 1.0:
+            logits_processor.append(TopPLogitsWarper(kwargs.pop("top_p")))
+        if self.kappa != 0:
+            logits_processor.append(self.logits_processor)
+
+        with torch.no_grad():
+            self.logits_processor.reset(n_gram)
+            output = self.model.generate(
+                **tokd_input,
+                max_new_tokens=max_new_tokens,
+                do_sample=do_sample,
+                logits_processor=logits_processor,
+                pad_token_id=self.tokenizer.eos_token_id,
+                **kwargs
+                )
+        output = output[:,tokd_input["input_ids"].shape[-1]:].cpu()
+
+        return_dict = {}
+
+        if return_scores:
+            cumulative_token_count = self.get_cumulative_token_count(self.id, output, n_gram = n_gram, return_dense=False)
+            cumulative_token_count = vstack([i[0] for i in cumulative_token_count], format="csr")
+            q_score, _, _ = self.watermarking_fn.q(cumulative_token_count, k_p = [self.k_p], use_tqdm=False)
+            return_dict["q_score"] = q_score[:,0]
+
+        if return_tokens:
+            return_dict["tokens"] = output
+
+        if return_text:
+            decoded_output = self.tokenizer.batch_decode(output, skip_special_tokens=True)
+            decoded_output = [i.strip() for i in decoded_output]
+            return_dict["text"] = decoded_output
+
+        if len(output) == 1:
+            for k, v in return_dict.items():
+                return_dict[k] = v[0]
+
+        if return_text and len(return_dict) == 0:
+            return decoded_output
+
+        return return_dict
+
+    def get_cumulative_token_count(
+            self,
+            ids : List[int] | int,
+            all_tokens : List[torch.Tensor] | torch.Tensor | List[np.ndarray] | np.ndarray | List[List[int]] | List[int],
+            n_gram : int = 2,
+            return_unshuffled_indices : bool = False,
+            use_tqdm : bool = False,
+            return_dense : bool = True,
+            batch_size : int = 2**8,
+            ) -> List[csr_matrix] | List[np.ndarray] | Tuple[List[csr_matrix], List[List[np.ndarray]]] | Tuple[List[np.ndarray], List[List[np.ndarray]]]:
+        if isinstance(ids, int):
+            ids = [ids]
+        if isinstance(all_tokens[0], int) or (isinstance(all_tokens, (np.ndarray, torch.Tensor)) and all_tokens.ndim == 1):
+            all_tokens = [all_tokens]
+        all_tokens = list(map(lambda x: x.cpu().numpy() if isinstance(x, torch.Tensor) else x, all_tokens))
+        max_length = max(map(len, all_tokens))
+        window = n_gram - 1
+
+        # Collect all unique seeds for psuedo-random number generation
+        key_index_dict = defaultdict(set)
+        all_keys = []
+        for i, tokens in enumerate(tqdm(all_tokens, desc="Collecting unique n-grams", disable=not use_tqdm)):
+            all_keys.append([])
+            for j in range(window, len(tokens)):
+                prev_token = tuple(tokens[j-window:j])
+                t = tokens[j]
+                if t >= self.N:
+                    break
+                key_index_dict[prev_token].add(t)
+                all_keys[i].append((prev_token, t))
+        key_index_dict = {k:tuple(v) for k,v in key_index_dict.items()}
+
+        use_mp = len(all_tokens) > batch_size * 4
+        if use_mp:
+            p = Pool(len(os.sched_getaffinity(0))-1)
+            pool_map = partial(p.imap, chunksize=batch_size)
+        else:
+            pool_map = map
+
+        # Generate permutations for all unique seeds
+        permutations = pool_map(
+            partial(self.logits_processor.permute.get_unshuffled_indices, ids),
+            key_index_dict.items())
+        permutations = tqdm(permutations, total=len(key_index_dict), desc="Getting permutations", disable=not use_tqdm)
+        for k, value in zip(key_index_dict.keys(), permutations):
+            key_index_dict[k] = value
+
+        # Assign indices to unshuffled_indices
+        unshuffled_indices: List[np.ndarray] = []  # [text x id x length]
+        for keys in tqdm(all_keys, desc="Assigning indices", disable=not use_tqdm):
+            if len(keys) == 0:
+                unshuffled_indices.append(np.zeros((len(ids), 0), dtype=np.min_scalar_type(self.N)))
+            else:
+                unshuffled_indices.append(np.stack([key_index_dict[key][t] for key, t in keys]).T)  # [id x length]
+
+        # Convert indices to counts
+        cumulative_token_count = pool_map(
+            partial(indices_to_counts, self.N, np.min_scalar_type(max_length)),
+            unshuffled_indices
+            )
+        cumulative_token_count = list(tqdm(cumulative_token_count, total=len(unshuffled_indices), desc="Counting tokens", disable=not use_tqdm))
+
+        if use_mp:
+            p.close()
+            p.join()
+
+        if return_dense:
+            cumulative_token_count = list(map(lambda x: x.toarray(), cumulative_token_count))
+
+        if return_unshuffled_indices:
+            return cumulative_token_count, unshuffled_indices
+        return cumulative_token_count
+
+    def verify(
+            self,
+            text : str | List[str],
+            id: Optional[int | List[int]] = None,
+            k_p : Optional[int | List[int]] = None,
+            return_ranking : bool = False,
+            return_extracted_k_p : bool = False,
+            return_counts : bool = False,
+            return_unshuffled_indices : bool = False,
+            use_tqdm : bool = False,
+            batch_size : int = 2**8,
+            ) -> np.ndarray | dict:
+        begin_time = time.time()
+
+        if id is None:
+            id = self.id
+
+        if isinstance(text, str):
+            texts = [text]
+        else:
+            texts = text
+
+        tokens = [np.array(self.tokenizer.encode(text, add_special_tokens=False), dtype=np.uint32) for text in tqdm(texts, desc="Tokenizing", disable=not use_tqdm)]
+
+        if isinstance(id, int):
+            ids = [id]
+        else:
+            ids = id
+
+        if isinstance(k_p, int):
+            k_ps = [k_p]
+        else:
+            k_ps = k_p
+
+        # Get cummulative token counts
+        start_time = time.time()
+        results = self.get_cumulative_token_count(ids, tokens, self.n_gram, return_unshuffled_indices, use_tqdm=use_tqdm, return_dense=False, batch_size=batch_size)
+        gc.collect()
+        if return_unshuffled_indices:
+            results, unshuffled_indices = results
+        results = vstack(results, format="csr")
+        if use_tqdm:
+            tqdm.write(f"Cummulative token counts done in {time.time() - start_time:.2f} seconds")
+
+        # Calculate Q score via dot product
+        start_time = time.time()
+        q_score, ranking, k_p_extracted = self.watermarking_fn.q(results, k_p = k_ps, batch = batch_size, use_tqdm = use_tqdm)
+        q_score, ranking = [i.reshape(-1, len(ids), i.shape[-1]) for i in (q_score, ranking)]  # [text x ids x k_p for i in (score, rank)]
+        k_p_extracted = k_p_extracted.reshape(-1, len(ids))  # [text x ids]
+        if use_tqdm:
+            tqdm.write(f"Q score calculated in {time.time() - start_time:.2f} seconds")
+
+        res = q_score # [text x ids x k_p]
+
+        if return_ranking or return_extracted_k_p or return_counts or return_unshuffled_indices:
+            res = {
+                "q_score": q_score,                             # [text x ids x k_p]
+                }
+            if return_ranking:
+                res["ranking"] = ranking                        # [text x ids x k_p]
+            if return_extracted_k_p:
+                res["k_p_extracted"] = k_p_extracted            # [text x ids]
+            if return_counts:
+                res["counts"] = results                         # [text x ids x k_p]
+            if return_unshuffled_indices:
+                res["unshuffled_indices"] = unshuffled_indices  # [text x ids x length]
+
+        if use_tqdm:
+            tqdm.write(f"Total time taken for verify: {time.time() - begin_time:.2f} seconds")
+
+        return res

waterfall/WatermarkingFn.py

@@ -0,0 +1,54 @@
+import numpy as np
+from tqdm import tqdm
+from multiprocessing import Pool
+import os
+from functools import partial
+from typing import List, Tuple
+from scipy.sparse import spmatrix
+
+class WatermarkingFn:
+    def __init__(self, id : int = 0, k_p : int = 1, N : int = 32000, kappa : float = 1.) -> None:
+        self.id = id
+        self.k_p = k_p
+        self.N = N
+        self.kappa = kappa
+        self.phi = None
+        self.dtype = np.min_scalar_type(self.N)
+
+    def _q(self, bins : np.ndarray | spmatrix, k_p : List[int]) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        raise NotImplementedError
+
+    def q(self,
+          bins : np.ndarray | spmatrix,
+          k_p : List[int],   # If set, only return the k_p-th element of the dot product and its ranking
+          batch : int = 2**8,
+          use_tqdm : bool = False,
+          ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        if bins.ndim == 1:
+            bins = bins[None,:]
+        res = []
+        bins_sum = bins.sum(axis=1).reshape(-1,1)
+        bins_sum[bins_sum == 0] = 1
+        batch_range = range(0, bins.shape[0], batch)
+        batched = (bins[i:i+batch] / bins_sum[i:i+batch] for i in batch_range)
+        use_mp = len(batch_range) > 4
+        if use_mp:
+            p = Pool(len(os.sched_getaffinity(0))-1)
+            pool_map = p.imap
+        else:
+            pool_map = map
+        res = pool_map(partial(self._q, k_p=k_p), batched)
+        if use_tqdm:
+            res_ = []
+            with tqdm(total=bins.shape[0], desc="Calculating dot product") as pbar:
+                for r in res:
+                    res_.append(r)
+                    pbar.update(len(r[0]))
+            res = res_
+        else:
+            res = list(res)
+        if use_mp:
+            p.close()
+            p.join()
+        k_p_strength, k_p_ranking, k_p_extracted = list(map(np.concatenate, zip(*res)))
+        return k_p_strength, k_p_ranking, k_p_extracted

waterfall/WatermarkingFnFourier.py

@@ -0,0 +1,32 @@
+from waterfall.WatermarkingFn import *
+import numpy as np
+from scipy.fft import rfft
+from scipy.sparse import isspmatrix
+
+class WatermarkingFnFourier(WatermarkingFn):
+    def __init__(self, id : int = 0, k_p : int = 1, N : int = 32000, kappa : float = 1.) -> None:
+        super().__init__(id = id, k_p = k_p, N = N, kappa = kappa)
+
+        freq = self.k_p
+        assert (freq > 0) and (freq < self.N), f"k_p must be 0<k_p<{self.N}, value provided is {freq}"
+
+        half_N = int(self.N/2)
+        if freq <= half_N:
+            self.phi = np.cos(np.arange(self.N)/self.N*2*np.pi*freq)
+        else:
+            freq -= half_N
+            self.phi = np.sin(np.arange(self.N)/self.N*2*np.pi*freq)
+        self.phi *= self.kappa
+
+        self.scaling_factor = 1
+
+    def _q(self, bins : np.ndarray | spmatrix, k_p : List[int]) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        if isspmatrix(bins):
+            bins = bins.todense()
+        q = rfft(bins, axis=-1)[:,1:-1].astype(np.complex64)
+        q = np.concatenate((np.real(q), np.imag(q)), axis=1)
+        q *= self.scaling_factor
+        k_p_strength = q[:,np.array(k_p)-1]
+        k_p_ranking = ((q[...,None,:] > k_p_strength[...,None]).sum(axis=-1)).astype(self.dtype)
+        k_p_extracted = (np.argmax(q, axis=-1) + 1).astype(self.dtype)
+        return k_p_strength, k_p_ranking, k_p_extracted

waterfall/WatermarkingFnSquare.py

@@ -0,0 +1,42 @@
+from waterfall.WatermarkingFn import *
+import numpy as np
+
+class WatermarkingFnSquare(WatermarkingFn):
+    def __init__(self, id : int = 0, k_p : int = 1, N : int = 32000, kappa : float = 1.) -> None:
+        super().__init__(id = id, k_p = k_p, N = N, kappa = kappa)
+
+        self.k_N = 0
+
+        if (self.N%2)==1:
+            self.k_N = 1
+            self.phi = np.ones(self.N) * self.kappa
+            self.phi[self.N//2:] *= -1
+            self.phis = [self.phi]
+            return
+
+        N = self.N
+        while not (N & 0b1):
+            N >>= 1
+            self.k_N += 1
+        assert (k_p > 0) and (k_p < (self.k_N * 2)), f"k_p {k_p} larger than available number of fns {self.k_N*2-1}"
+
+        self.phis = np.empty((self.k_N*2-1, self.N), dtype=np.float32)
+        for i in range(self.k_N*2-1):
+            k_p = i+1
+            if k_p <= self.k_N:
+                self.phis[i] = (-1)**(np.floor(np.arange(self.N)*2**k_p/self.N)) * self.kappa
+            else:
+                k_p -= self.k_N
+                self.phis[i] = (-1)**(np.floor(np.arange(self.N)*2**k_p/self.N+0.5)) * self.kappa
+
+        self.phi = self.phis[self.k_p-1]
+
+        self.scaling_factor = 1 / self.kappa
+
+    def _q(self, bins : np.ndarray | spmatrix, k_p : List[int]) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        q = bins.dot(self.phis.T)
+        q *= self.scaling_factor
+        k_p_strength = q[:,np.array(k_p)-1]
+        k_p_ranking = ((q[...,None,:] > k_p_strength[...,None]).sum(axis=-1)).astype(self.dtype)
+        k_p_extracted = (np.argmax(q, axis=-1) + 1).astype(self.dtype)
+        return k_p_strength, k_p_ranking, k_p_extracted

waterfall/permute.py

@@ -0,0 +1,67 @@
+import numpy as np
+import psutil
+from collections import OrderedDict
+import gc
+from typing import TypeVar, Tuple
+
+T = TypeVar('T')
+
+class LRUCache:
+    def __init__(self, capacity: int = 1000) -> None:
+        self.cache = OrderedDict()
+        self.capacity = capacity
+        self.cache_hits : int = 0
+        self.cache_misses : int = 0
+
+    def get(self, key: Tuple) -> T | None:
+        if key not in self.cache:
+            self.cache_misses += 1
+            return None
+        else:
+            self.cache_hits += 1
+            # Move the accessed item to the end of the OrderedDict to mark it as recently used.
+            self.cache.move_to_end(key)
+            return self.cache[key]
+
+    def __str__(self) -> str:
+        gc.collect()
+        return f"Cache hits: {self.cache_hits}, misses: {self.cache_misses}, rate: {self.cache_hits/(max(self.cache_hits+self.cache_misses, 1)):.2f}"
+
+    def put(self, key: Tuple, value: T) -> None:
+        if key in self.cache:
+            # Update the value and move it to the end.
+            self.cache.move_to_end(key)
+        self.cache[key] = value
+        if len(self.cache) > self.capacity:
+            # Remove the first key-value pair which is the least recently used.
+            del self.cache[next(iter(self.cache))]
+
+    def clear(self) -> None:
+        self.cache.clear()
+        gc.collect()
+
+class Permute:
+    permutations = LRUCache()
+    def __init__(self, N : int = 128000) -> None:
+        self.N = N
+        self.dtype = np.min_scalar_type(self.N)
+        assert self.dtype.kind == 'u', "N must be a positive integer"
+        size_per_permutation_in_bytes = N * self.dtype.itemsize
+        cache_size = int(psutil.virtual_memory().total * 0.02 / size_per_permutation_in_bytes)  # 2% of total memory
+        self.permutations.capacity = cache_size
+
+    def get_permutation(self, prev_tok, id : int, cache : bool = False) -> np.ndarray:
+        key = (id, *prev_tok)
+        if cache:
+            permutation = self.permutations.get(key)
+            if permutation is None:
+                permutation = np.random.RandomState(key).permutation(self.N).astype(self.dtype)
+                self.permutations.put(key, permutation)
+        else:
+            permutation = np.random.RandomState(key).permutation(self.N).astype(self.dtype)
+        return permutation
+
+    def get_unshuffled_indices(self, ids, args) -> dict[int, np.ndarray]:
+        key, indices = args
+        permutation = np.stack([self.get_permutation(key, id) for id in ids])
+        return {k: v for k, v in zip(indices, permutation[:,indices].T)}

waterfall/watermark.py

@@ -0,0 +1,307 @@
+import argparse
+import logging
+import os
+import torch
+from typing import List, Literal, Optional, Tuple
+
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from sentence_transformers import SentenceTransformer
+from tqdm.auto import tqdm
+
+from waterfall.WatermarkingFnFourier import WatermarkingFnFourier
+from waterfall.WatermarkingFnSquare import WatermarkingFnSquare
+from waterfall.WatermarkerBase import Watermarker
+
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+PROMPT = (
+    "Paraphrase the user provided text while preserving semantic similarity. "
+    "Do not include any other sentences in the response, such as explanations of the paraphrasing. "
+    "Do not summarize."
+)
+PRE_PARAPHRASED = "Here is a paraphrased version of the text while preserving the semantic similarity:\n\n"
+
+def detect_gpu() -> str:
+    """
+    Use torch to detect if MPS, CUDA, or neither (default CPU)
+    are available.
+
+    Returns:
+        String for the torch device available.
+    """
+    if torch.backends.mps.is_available():
+        return "mps"
+    elif torch.cuda.is_available():
+        return 'cuda'
+    else:
+        return 'cpu'
+
+def watermark(
+    T_o: str,
+    watermarker: Watermarker,
+    sts_model: SentenceTransformer,
+    num_beam_groups: int = 4,
+    beams_per_group: int = 2,
+    STS_scale:float = 2.0,
+    diversity_penalty: float = 0.5,
+    max_new_tokens: Optional[int] = None,
+) -> str:
+    paraphrasing_prompt = watermarker.tokenizer.apply_chat_template(
+        [
+            {"role":"system", "content":PROMPT},
+            {"role":"user", "content":T_o},
+        ], tokenize=False, add_generation_prompt = True) + PRE_PARAPHRASED
+
+    watermarked = watermarker.generate(
+        paraphrasing_prompt,
+        return_scores = True,
+        max_new_tokens = int(len(paraphrasing_prompt) * 1.5) if max_new_tokens is None else max_new_tokens,
+        do_sample = False, temperature=None, top_p=None,
+        num_beams = num_beam_groups * beams_per_group,
+        num_beam_groups = num_beam_groups,
+        num_return_sequences = num_beam_groups * beams_per_group,
+        diversity_penalty = diversity_penalty,
+        )
+
+    # Select best paraphrasing based on q_score and semantic similarity
+    sts_scores = STS_scorer(T_o, watermarked["text"], sts_model)
+    selection_score = sts_scores * STS_scale + torch.from_numpy(watermarked["q_score"])
+    selection = torch.argmax(selection_score)
+
+    T_w = watermarked["text"][selection]
+
+    return T_w
+
+def verify_texts(texts: List[str], id: int, 
+                     watermarker: Optional[Watermarker] = None, 
+                     k_p: Optional[int] = None, 
+                     model_path: Optional[str] = "meta-llama/Llama-3.1-8B-Instruct"
+                     ) -> Tuple[float,float]:
+    """Returns the q_score and extracted k_p"""
+
+    if watermarker is None:
+        assert model_path is not None, "model_path must be provided if watermarker is not passed"
+        tokenizer = AutoTokenizer.from_pretrained(model_path)
+        watermarker = Watermarker(tokenizer=tokenizer)
+    
+    if k_p is None:
+        k_p = watermarker.k_p
+
+    verify_results = watermarker.verify(texts, id=[id], k_p=[k_p], return_extracted_k_p=True)  # results are [text x id x k_p]
+    q_score = verify_results["q_score"]
+    k_p_extracted = verify_results["k_p_extracted"]
+
+    return q_score[:,0,0], k_p_extracted[:, 0]
+
+def STS_scorer_batch(
+    original_texts: List[str],
+    test_texts: List[List[str]],
+    sts_model: SentenceTransformer
+) -> torch.Tensor:
+
+    assert len(original_texts) == len(test_texts), "original_texts and test_texts must have the same length"
+    assert all(len(test_texts[0]) == len(sublist) for sublist in test_texts[1:]), "All sublists in test_texts must have the same length"
+
+    all_text = original_texts + [text for sublist in test_texts for text in sublist]
+    embeddings = sts_model.encode(all_text, convert_to_tensor=True, normalize_embeddings=True)
+    original_embeddings = embeddings[:len(original_texts)]
+    test_embeddings = embeddings[len(original_texts):].reshape(len(test_texts), -1, embeddings.shape[1])
+    cos_sim = torch.einsum('ik,ijk->ij', original_embeddings, test_embeddings).cpu()
+    return cos_sim
+
+def STS_scorer(
+    original_text: str,
+    test_texts: str | List[str],
+    sts_model: SentenceTransformer
+) -> float | torch.Tensor:
+    cos_sim = STS_scorer_batch(
+        original_texts=[original_text],
+        test_texts=[[test_texts] if isinstance(test_texts, str) else test_texts],
+        sts_model=sts_model
+    )[0]
+    if isinstance(test_texts, str):
+        cos_sim = cos_sim.item()
+    return cos_sim
+
+def watermark_texts(
+    T_os: List[str],
+    id: Optional[int] = None,
+    k_p: int = 1,
+    kappa: float = 2.0,
+    model_path: str = "meta-llama/Llama-3.1-8B-Instruct",
+    torch_dtype: torch.dtype = torch.bfloat16,
+    sts_model_path: str = "sentence-transformers/all-mpnet-base-v2",
+    watermark_fn: Literal["fourier", "square"] = "fourier",
+    watermarker: Optional[Watermarker] = None,
+    sts_model: Optional[SentenceTransformer] = None,
+    device: str = detect_gpu(),
+    num_beam_groups: int = 4,
+    beams_per_group: int = 2,
+    diversity_penalty: float = 0.5,
+    STS_scale:float = 2.0,
+    use_tqdm: bool = False,
+) -> List[str]:
+    if watermark_fn == 'fourier':
+        watermarkingFnClass = WatermarkingFnFourier
+    elif watermark_fn == 'square':
+        watermarkingFnClass = WatermarkingFnSquare
+    else:
+        raise ValueError("Invalid watermarking function")
+
+    if watermarker is None:
+        assert model_path is not None, "model_path must be provided if watermarker is not passed"
+        model = AutoModelForCausalLM.from_pretrained(
+            model_path,
+            torch_dtype=torch_dtype,
+            device_map=device,
+            )
+        tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+        watermarker = Watermarker(tokenizer=tokenizer, model=model, id=id, kappa=kappa, k_p=k_p, watermarkingFnClass=watermarkingFnClass)
+    else:
+        tokenizer = watermarker.tokenizer
+        device = watermarker.model.device
+        id = watermarker.id
+
+    if id is None:
+        raise Exception("ID or Watermarker class must be passed to watermark_texts.")
+
+    if sts_model is None:
+        assert sts_model_path is not None, "sts_model_path must be provided if sts_model is not passed"
+        sts_model = SentenceTransformer(sts_model_path, device=device)
+
+    T_ws = []
+
+    for T_o in tqdm(T_os, desc="Watermarking texts",  disable=not use_tqdm):
+        T_w = watermark(
+            T_o,
+            watermarker = watermarker,
+            sts_model = sts_model,
+            num_beam_groups = num_beam_groups,
+            beams_per_group = beams_per_group,
+            diversity_penalty = diversity_penalty,
+            STS_scale = STS_scale,
+            )
+        T_ws.append(T_w)
+
+    return T_ws
+
+def pretty_print(
+        T_o: str, T_w: str,
+        sts_score: float,
+        T_o_q_score: float, T_w_q_score: float,
+        k_p: int, T_w_k_p: int,
+        ) -> None:
+    print(f"\nOriginal text T_o:\n\n{T_o}\n")
+    print(f"\nWatermarked text T_w:\n\n{T_w}\n")
+
+    # Original text
+    print(f"Verification score of T_o: \033[93m{T_o_q_score:.4f}\033[0m")
+
+    # Watermarked text
+    print(f"Verification score of T_w: \033[92m{T_w_q_score:.4f}\033[0m\n")
+
+    print(f"STS score of T_w         : \033[94m{sts_score:.4f}\033[0m\n")
+
+    # Extract from watermarked text
+    print(f"Watermarking k_p         : \033[95m{k_p}\033[0m")
+    print(f"Extracted k_p from T_w   : \033[96m{T_w_k_p}\033[0m\n")
+
+def main():
+    parser = argparse.ArgumentParser(description='generate text watermarked with a key')
+    parser.add_argument('--id',default=42,type=int,
+        help='id: unique ID')
+    parser.add_argument('--kappa',default=2.,type=float,
+        help='kappa: watermarking strength')
+    parser.add_argument('--k_p', default=1, type=int,
+        help="k_p: Perturbation key")
+    parser.add_argument('--model', default='meta-llama/Llama-3.1-8B-Instruct', type=str,
+        help="watermarking model")
+    parser.add_argument('--sts_model', default='sentence-transformers/all-mpnet-base-v2', type=str,
+        help="STS model")
+    parser.add_argument('--T_o', default=None, type=str,
+        help="original_text")
+    parser.add_argument('--watermark_fn', default='fourier', type=str,
+        help="watermarking function, can be 'fourier' or 'square'")
+    parser.add_argument('--device', default=detect_gpu(), type=str,
+        help="device to use for generation")
+    parser.add_argument('--num_beam_groups', default=4, type=int,
+        help="number of beam groups for generation")
+    parser.add_argument('--beams_per_group', default=2, type=int,
+        help="number of beams per group for generation")
+    parser.add_argument('--diversity_penalty', default=0.5, type=float,
+        help="diversity penalty for group beam search")
+    parser.add_argument('--STS_scale', default=2.0, type=float,
+        help="scale factor for trade-off between STS and q score. Higher means more emphasis on STS.")
+
+    args = parser.parse_args()
+
+    if args.watermark_fn == 'fourier':
+        watermarkingFnClass = WatermarkingFnFourier
+    elif args.watermark_fn == 'square':
+        watermarkingFnClass = WatermarkingFnSquare
+    else:
+        # Add any other self-defined watermarking functions here
+        raise ValueError("Invalid watermarking function")
+
+    id = args.id
+    kappa = args.kappa
+    k_p = args.k_p
+    model_name_or_path = args.model
+    sts_model_name = args.sts_model
+    T_o = args.T_o
+    device = args.device
+    num_beam_groups = args.num_beam_groups
+    beams_per_group = args.beams_per_group
+    diversity_penalty = args.diversity_penalty
+    STS_scale = args.STS_scale
+
+    if args.T_o is None:
+        T_o = "Protecting intellectual property (IP) of text such as articles and code is increasingly important, especially as sophisticated attacks become possible, such as paraphrasing by large language models (LLMs) or even unauthorized training of LLMs on copyrighted text to infringe such IP. However, existing text watermarking methods are not robust enough against such attacks nor scalable to millions of users for practical implementation."
+    T_os = [T_o]    # Replace with your own list of texts to watermark
+
+    # Initialize tokenizer and model
+    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_name_or_path,
+        torch_dtype=torch.bfloat16,
+        device_map=device,
+        )
+
+    watermarker = Watermarker(tokenizer=tokenizer, model=model, id=id, kappa=kappa, k_p=k_p, watermarkingFnClass=watermarkingFnClass)
+
+    sts_model = SentenceTransformer(sts_model_name, device=device)
+
+    T_ws = watermark_texts(
+        T_os, id, k_p, kappa,
+        watermarker=watermarker, sts_model=sts_model,
+        beams_per_group=beams_per_group,
+        num_beam_groups=num_beam_groups,
+        diversity_penalty=diversity_penalty,
+        STS_scale=STS_scale,
+        use_tqdm=True
+        )
+
+    # watermarker = Watermarker(tokenizer=tokenizer, model=None, id=id, k_p=k_p, watermarkingFnClass=watermarkingFnClass)   # If only verifying the watermark, do not need to instantiate the model
+    q_scores, extracted_k_ps = verify_texts(T_os + T_ws, id, watermarker, k_p=k_p)
+
+    for i in range(len(T_os)):
+        # Handle the case where this is being run
+        # in an IDE or something else without terminal size
+        try:
+            column_size = os.get_terminal_size().columns
+        except OSError as ose:
+            column_size = 80
+
+        print("=" * column_size)
+
+        sts_score = STS_scorer(T_os[i], T_ws[i], sts_model)
+        pretty_print(
+            T_os[i], T_ws[i],
+            sts_score,
+            q_scores[i], q_scores[i + len(T_os)],
+            k_p, extracted_k_ps[i + len(T_os)],
+        )
+
+if __name__ == "__main__":
+    main()

waterfall-0.1.0.dist-info/METADATA

@@ -0,0 +1,158 @@
+Metadata-Version: 2.4
+Name: waterfall
+Version: 0.1.0
+Summary: Scalable Framework for Robust Text Watermarking and Provenance for LLMs
+Author-email: Xinyuan Niu <aperture@outlook.sg>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/aoi3142/Waterfall
+Project-URL: Issues, https://github.com/aoi3142/Waterfall/issues
+Classifier: Programming Language :: Python :: 3
+Classifier: Operating System :: OS Independent
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: accelerate>=0.29.0
+Requires-Dist: numpy>=2.0.0
+Requires-Dist: scipy>=1.13.0
+Requires-Dist: sentence-transformers>=3.0.0
+Requires-Dist: torch>=2.3.0
+Requires-Dist: transformers>=4.43.1
+Dynamic: license-file
+
+# Waterfall: Scalable Framework for Robust Text Watermarking and Provenance for LLMs [EMNLP 2024 Main Long]
+Gregory Kang Ruey Lau*, Xinyuan Niu*, Hieu Dao, Jiangwei Chen, Chuan-Sheng Foo, Bryan Kian Hsiang Low
+
+[EMNLP](https://aclanthology.org/2024.emnlp-main.1138/) | [ArXiv](https://arxiv.org/abs/2407.04411) | [PDF](https://arxiv.org/pdf/2407.04411)
+
+## TL;DR: Training-free framework for text watermarking that is scalable, robust to LLM attacks, and applicable to original text of multiple types 
+
+![Alt text](Images/Problem_formulation.jpg "")
+
+1. Watermark original text $T_o$ with watermark key $\mu$ → watermarked text $T_w$ with same semantic content
+
+2. Adversaries try to claim IP by plagiarizing text (e.g. paraphrasing), or by using text to train their own LLMs without authorization
+
+3. Clients can quickly verify whether a suspected text $T_{sus}$ contains the watermark and originated from $T_o$
+
+Note: This code has been slightly modified from the implementation of the experiments in the paper. Refer to Appendix L.6 for details.
+
+# Abstract
+Protecting intellectual property (IP) of text such as articles and code is increasingly important, especially as sophisticated attacks become possible, such as paraphrasing by large language models (LLMs) or even unauthorized training of LLMs on copyrighted text to infringe such IP. However, existing text watermarking methods are not robust enough against such attacks nor scalable to millions of users for practical implementation. In this paper, we propose Waterfall, the first training-free framework for robust and scalable text watermarking applicable across multiple text types (e.g., articles, code) and languages supportable by LLMs, for general text and LLM data provenance. Waterfall comprises several key innovations, such as being the first to use LLM as paraphrasers for watermarking along with a novel combination of techniques that are surprisingly effective in achieving robust verifiability and scalability. We empirically demonstrate that Waterfall achieves significantly better scalability, robust verifiability, and computational efficiency compared to SOTA article-text watermarking methods, and also showed how it could be directly applied to the watermarking of code.
+
+# Watermark process
+
+![Alt text](Images/Watermarking_process.png "")
+
+1. Original text $T_o$ is fed into LLM paraphraser to produce initial logits $L$.
+
+2. Unique ID $\mu$ and preceding $n-1$ tokens form the permutation key $k_\pi$ which seed a pseudo-random permutation which permutes the initial logits from $V_o$ space into $V_w$ space.
+
+3. Perturbation key $k_p$ selects a perturbation function $\mathcal{F}_1$ out of a family of orthogonal functions. $\mathcal{F}_1$ is added to the permuted logits.
+
+4. The perturbed logits are permuted back from $V_w$ space into $V_o$ space with the inverse of the permutation in step 2.
+
+5. A token is sampled from the perturbed logits $\check{L}$ and is appended to the watermarked text.
+
+6. Append the generated token to the prompt and continue autoregressive generation (steps 1-5) until the eos token.
+
+# Verification of un/watermarked text
+
+![Alt text](Images/Illustration.gif "Text watermarked with a sine-watermark shows the watermark signal when verified with the correct key")
+
+1. For each token $\hat{w}$ in the watermarked text $T_w$ (original text is not required), use the unique ID $\mu$ and preceding $n-1$ tokens to permute the token index of $\hat{w}$ from $V_o$ space into $V_w$ space.
+
+2. Count the tokens to get a cumulative token distribution $C$ in $V_w$ space.
+
+3. Calculate the watermark score $q$ by taking the inner product of the cumulative token distribution $C$ with the perturbation function $\mathcal{F}_1$.
+
+4. Watermarked text will have $C$ that resembles $\mathcal{F}_1$, resulting in high $q$. Unwatermarked text or text watermarked with different ID $\mu$ will have a flat $C$, and text watermarked with different $k_p$ will have a $C$ that is orthogonal to $\mathcal{F}_1$, resulting in low watermark score $q$.
+
+# Extraction of perturbation key $k_p$
+
+1. Perform steps 1-2 in verification.
+
+2. Calculate the watermark scores $q$ for all perturbation functions $\mathcal{F}_1$ in family of orthogonal functions.
+
+3. Extracted $k_p$ corresponds to the perturbation function $\mathcal{F}_1$ with the highest scoring watermark score in step 2.
+
+# Using our code
+
+[Optional]
+If using `conda` (or other pkg managers), it is highly advisable to create a new environment
+
+```sh
+conda create -n waterfall python=3.11 --yes   `# Compatible with python version higher than 3.10`
+conda activate waterfall
+```
+
+Clone and install our package
+```sh
+git clone https://github.com/aoi3142/Waterfall.git
+pip install -e Waterfall        `# Install in 'editable' mode with '-e', can be omitted`
+```
+
+## Minimal demonstration for Waterfall watermarking
+
+Use the command `waterfall_demo` to watermark a piece of text, and then verify the presence of the watermark in the watermarked text
+```sh
+waterfall_demo
+```
+
+Additional arguments
+```sh
+waterfall_demo \
+  --T_o          "TEXT TO WATERMARK"              `# Original text to watermark`  \
+  --id           42                               `# Unique watermarking ID`      \
+  --k_p          1                                `# Additional perturbation key` \
+  --kappa        2                                `# Watermark strength`          \
+  --model        meta-llama/Llama-3.1-8B-Instruct `# Paraphrasing LLM`            \
+  --watermark_fn fourier                          `# fourier/square watermark`    \
+  --device       cuda                             `# Use cuda/cpu`
+```
+
+## Using our code to watermark and verify
+
+To watermark texts
+
+```python
+from waterfall.watermark import watermark_texts
+
+id = 1                        # specify your watermarking ID
+texts = ["...", "..."]        # Assign texts to be watermarked
+
+watermarked_text = watermark_texts(texts, id)         # List of strings
+```
+
+To verify watermark strength of texts
+
+```python
+from waterfall.watermark import verify_texts
+
+id = 1                        # specify your watermarking ID
+test_texts = ["...", "..."]   # Suspected texts to verify
+
+watermark_strength = verify_texts(test_texts, id)[0]   # np array of floats
+```
+
+# Code structure
+
+- `watermark.py`              : Sample watermarking script used by with `watermark_demo` command, includes beam search and other optimizations
+- `WatermarkerBase.py`        : Underlying generation and verification code provided by `Watermarker` class
+- `WatermarkingFn.py`         : Abstract class `WatermarkingFn` for watermarking functions, inherit it to create new perturbation functions
+- `WatermarkingFnFourier.py`  : Fourier watermarking function `WatermarkingFnFourier` inherited from `WatermarkingFn`
+- `WatermarkingFnSquare.py`   : Square watermarking function `WatermarkingFnSquare` inherited from `WatermarkingFn`
+
+# BibTeX
+```
+@inproceedings{lau2024waterfall,
+  title={Waterfall: Scalable Framework for Robust Text Watermarking and Provenance for {LLM}s},
+  author={Lau, Gregory Kang Ruey and Niu, Xinyuan and Dao, Hieu and Chen, Jiangwei and Foo, Chuan-Sheng and Low, Bryan Kian Hsiang},
+  booktitle={Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing},
+  year={2024},
+  month={nov},
+  address={Miami, Florida, USA},
+  url={https://aclanthology.org/2024.emnlp-main.1138/},
+  doi={10.18653/v1/2024.emnlp-main.1138},
+  pages={20432--20466},
+}
+```

waterfall-0.1.0.dist-info/RECORD

@@ -0,0 +1,13 @@
+waterfall/WatermarkerBase.py,sha256=ou78I1XisalHbJLqyST6ryuLjtkFnY7Y60fUKdIwLy4,12905
+waterfall/WatermarkingFn.py,sha256=-b-kGRdL0a7eKRqJmcHPAR_rCjxQYnsg1Ne6bTwBc1I,1931
+waterfall/WatermarkingFnFourier.py,sha256=QYayAQYwi1dQkDIyqmvhU568VhrVYTVy47HkI8F8SZs,1358
+waterfall/WatermarkingFnSquare.py,sha256=2PAO05DdKT02npo7GDf_82D520nP7kGAWK6H4E4JMt4,1638
+waterfall/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+waterfall/permute.py,sha256=RwxOHFhx_VSOhhFwy5s79YgwTUBkfW2-LCCXYR3VT2o,2582
+waterfall/watermark.py,sha256=whiNhPwWNNIZwXMH6r7QzEE3A7Niq2Ro9elA1iSRoxI,11952
+waterfall-0.1.0.dist-info/licenses/LICENSE,sha256=zAtaO-k41Q-Q4Etl4bzuh7pgNJsPH-dYfzvznRa0OvM,11341
+waterfall-0.1.0.dist-info/METADATA,sha256=ONpoos0Pyx43h4ntb9Fs4l78F3GZNxq__Ox05GMFD8A,8221
+waterfall-0.1.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+waterfall-0.1.0.dist-info/entry_points.txt,sha256=XXnUzuWXu2nc9j4WAll9tq6HyodN_8WJLjeG0O4Y2Gw,60
+waterfall-0.1.0.dist-info/top_level.txt,sha256=5rTgijeT9V5GRCwIDZmhjeZ4khgH1lmfhS9ZmdUUCKQ,10
+waterfall-0.1.0.dist-info/RECORD,,

waterfall-0.1.0.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (80.9.0)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

waterfall-0.1.0.dist-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+waterfall_demo = waterfall.watermark:main

waterfall-0.1.0.dist-info/licenses/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2024 Xinyuan Niu
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

waterfall-0.1.0.dist-info/top_level.txt

@@ -0,0 +1 @@
+waterfall