synth-ai 0.2.4.dev5__py3-none-any.whl → 0.2.4.dev7__py3-none-any.whl

synth_ai/environments/tasks/filters.py

@@ -1,6 +1,8 @@
-from typing import Any, Collection, Optional
+from collections.abc import Collection
 from dataclasses import dataclass
-from synth_ai.environments.tasks.core import TaskInstanceMetadataFilter, TaskInstance
+from typing import Any
+
+from synth_ai.environments.tasks.core import TaskInstance, TaskInstanceMetadataFilter
 
 
 @dataclass
@@ -18,8 +20,8 @@ class ValueFilter(TaskInstanceMetadataFilter):
 @dataclass
 class RangeFilter(TaskInstanceMetadataFilter):
     key: str
-    min_val: Optional[float] = None
-    max_val: Optional[float] = None
+    min_val: float | None = None
+    max_val: float | None = None
 
     def __call__(self, instance: TaskInstance) -> bool:
         instance_value = getattr(instance.metadata, self.key, None)

synth_ai/environments/tasks/utils.py

@@ -2,17 +2,19 @@
 Utility functions and generic filters for taskset creation.
 """
 
-from typing import Any, Collection, Optional, List, Set
+from collections.abc import Collection
+from typing import Any
 from uuid import UUID, uuid4
+
 from synth_ai.environments.tasks.core import (
-    TaskInstanceMetadataFilter,
-    TaskInstanceSet,
     SplitInfo,
     TaskInstance,
+    TaskInstanceMetadataFilter,
+    TaskInstanceSet,
 )
 
 
-def parse_or_new_uuid(raw_id: Optional[str]) -> UUID:
+def parse_or_new_uuid(raw_id: str | None) -> UUID:
     """
     Parse a raw ID string into a UUID, or generate a new one if invalid or missing.
     """
@@ -43,8 +45,8 @@ class RangeFilter(TaskInstanceMetadataFilter):
     def __init__(
         self,
         key: str,
-        min_value: Optional[float] = None,
-        max_value: Optional[float] = None,
+        min_value: float | None = None,
+        max_value: float | None = None,
     ):
         self.key = key
         self.min_value = min_value
@@ -62,15 +64,15 @@ class RangeFilter(TaskInstanceMetadataFilter):
 def make_taskset(
     name: str,
     description: str,
-    instances: List[TaskInstance],
-    val_filter: Optional[TaskInstanceMetadataFilter] = None,
-    test_filter: Optional[TaskInstanceMetadataFilter] = None,
+    instances: list[TaskInstance],
+    val_filter: TaskInstanceMetadataFilter | None = None,
+    test_filter: TaskInstanceMetadataFilter | None = None,
 ) -> TaskInstanceSet:
     """
     Assemble a TaskInstanceSet by applying optional validation and test filters.
     """
-    val_ids: Set[Any] = set()
-    test_ids: Set[Any] = set()
+    val_ids: set[Any] = set()
+    test_ids: set[Any] = set()
     if val_filter:
         val_ids = {inst.id for inst in instances if val_filter(inst)}
     if test_filter:

synth_ai/evals/base.py

@@ -1,9 +1,8 @@
-from typing import List
 
 
 class Judgement:
     def __init__(
-        self, criteria: str, score: float, reasoning: str = "", evidence: List[str] = None
+        self, criteria: str, score: float, reasoning: str = "", evidence: list[str] = None
     ):
         self.criteria = criteria
         self.score = score
@@ -12,5 +11,5 @@ class Judgement:
 
 
 class BaseEval:
-    async def run(self, data: any) -> List[Judgement]:
+    async def run(self, data: any) -> list[Judgement]:
         pass

synth_ai/experimental/synth_oss.py

@@ -1,5 +1,5 @@
-
-"""
+# ruff: noqa
+'''
 Synth OSS Integration Module
 
 This module provides integration with Synth's open-source inference and training APIs
@@ -336,7 +336,7 @@ Implementation sketch (backend == "synth")
 The method is a *no-op* for the default (OpenAI) backend so existing code keeps
 working.
 
-"""
+'''
 
 
 """
@@ -443,4 +443,4 @@ async def warmup(
 So: **the existing endpoint does not yet support GPU selection; we need to add
 the small change above on the `learning_v2` side and then LM.warmup can request
 specific GPUs.**
-"""
+"""

synth_ai/learning/gateway.py

@@ -1,4 +1,2 @@
-
-
 class OfflineGateway:
-    pass
+    pass

synth_ai/learning/prompts/banking77_injection_eval.py

@@ -0,0 +1,168 @@
+"""
+Banking77 in-context injection evals (async, not tests)
+
+Samples a handful of Banking77 prompts and evaluates multiple override
+contexts in parallel, printing simple accuracy for each.
+
+Usage
+- Keys in .env (GROQ_API_KEY, etc.)
+- Run: uv run -q python -m synth_ai.learning.prompts.banking77_injection_eval
+  Optional env:
+    - N_SAMPLES=20 (default)
+    - MODEL=openai/gpt-oss-20b (default)
+    - VENDOR=groq (default)
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import random
+from typing import Any
+
+from datasets import load_dataset
+from dotenv import load_dotenv
+from synth_ai.lm.core.main_v3 import LM, build_messages
+from synth_ai.lm.overrides import LMOverridesContext
+
+
+async def classify_one(lm: LM, text: str, label_names: list[str]) -> str:
+    labels_joined = ", ".join(label_names)
+    system_message = (
+        "You are an intent classifier for the Banking77 dataset. "
+        "Given a customer message, respond with exactly one label from the list. "
+        "Return only the label text with no extra words.\n\n"
+        f"Valid labels: {labels_joined}"
+    )
+    user_message = f"Message: {text}\nLabel:"
+    messages = build_messages(system_message, user_message, images_bytes=None, model_name=lm.model)
+    resp = await lm.respond_async(messages=messages)
+    return (resp.raw_response or "").strip()
+
+
+def choose_label(pred: str, label_names: list[str]) -> str:
+    norm_pred = pred.strip().lower()
+    label_lookup = {ln.lower(): ln for ln in label_names}
+    mapped = label_lookup.get(norm_pred)
+    if mapped is not None:
+        return mapped
+
+    # Fallback: choose the label with the highest naive token overlap
+    def score(cand: str) -> int:
+        c = cand.lower()
+        return sum(1 for w in c.split() if w in norm_pred)
+
+    return max(label_names, key=score)
+
+
+async def eval_context(
+    lm: LM,
+    items: list[tuple[str, str]],
+    label_names: list[str],
+    ctx_name: str,
+    specs: list[dict[str, Any]],
+) -> tuple[str, int, int]:
+    correct = 0
+    with LMOverridesContext(specs):
+        tasks = [classify_one(lm, text, label_names) for text, _ in items]
+        results = await asyncio.gather(*tasks, return_exceptions=True)
+    for (text, gold), pred in zip(items, results, strict=False):
+        if isinstance(pred, Exception):
+            # Treat exceptions as incorrect
+            continue
+        mapped = choose_label(pred, label_names)
+        correct += int(mapped == gold)
+    return (ctx_name, correct, len(items))
+
+
+async def main() -> None:
+    load_dotenv()
+
+    n = int(os.getenv("N_SAMPLES", "20"))
+    model = os.getenv("MODEL", "openai/gpt-oss-20b")
+    vendor = os.getenv("VENDOR", "groq")
+
+    lm = LM(model=model, vendor=vendor, temperature=0.0)
+
+    print("Loading Banking77 dataset (split='test')...")
+    ds = load_dataset("banking77", split="test")
+    label_names: list[str] = ds.features["label"].names  # type: ignore
+
+    idxs = random.sample(range(len(ds)), k=min(n, len(ds)))
+    items = [
+        (ds[i]["text"], label_names[int(ds[i]["label"])])  # (text, gold_label)
+        for i in idxs
+    ]
+
+    # Define a few override contexts to compare
+    contexts: list[dict[str, Any]] = [
+        {
+            "name": "baseline (no overrides)",
+            "overrides": [],
+        },
+        {
+            "name": "nonsense prompt injection (expected worse)",
+            "overrides": [
+                {
+                    "match": {"contains": "", "role": "user"},
+                    "injection_rules": [
+                        # Heavily corrupt user text by replacing vowels
+                        {"find": "a", "replace": "x"},
+                        {"find": "e", "replace": "x"},
+                        {"find": "i", "replace": "x"},
+                        {"find": "o", "replace": "x"},
+                        {"find": "u", "replace": "x"},
+                        {"find": "A", "replace": "X"},
+                        {"find": "E", "replace": "X"},
+                        {"find": "I", "replace": "X"},
+                        {"find": "O", "replace": "X"},
+                        {"find": "U", "replace": "X"},
+                    ],
+                }
+            ],
+        },
+        {
+            "name": "injection: atm->ATM, txn->transaction",
+            "overrides": [
+                {
+                    "match": {"contains": "atm", "role": "user"},
+                    "injection_rules": [
+                        {"find": "atm", "replace": "ATM"},
+                        {"find": "txn", "replace": "transaction"},
+                    ],
+                }
+            ],
+        },
+        {
+            "name": "params: temperature=0.0",
+            "overrides": [
+                {"match": {"contains": ""}, "params": {"temperature": 0.0}},
+            ],
+        },
+        {
+            "name": "model override: 20b->120b",
+            "overrides": [
+                {"match": {"contains": ""}, "params": {"model": "openai/gpt-oss-120b"}},
+            ],
+        },
+    ]
+
+    print(f"\nEvaluating {len(contexts)} contexts on {len(items)} Banking77 samples (async)...")
+
+    # Evaluate each context sequentially but batched (each context classifies in parallel)
+    results: list[tuple[str, int, int]] = []
+    for ctx in contexts:
+        name = ctx["name"]
+        specs = ctx["overrides"]
+        print(f"Evaluating: {name} ...")
+        res = await eval_context(lm, items, label_names, name, specs)
+        results.append(res)
+
+    print("\nResults:")
+    for name, correct, total in results:
+        acc = correct / total if total else 0.0
+        print(f"- {name}: {correct}/{total} correct ({acc:.2%})")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

synth_ai/learning/prompts/hello_world_in_context_injection_ex.py

@@ -0,0 +1,213 @@
+"""
+Hello World: Banking77 intent classification with in-context injection
+
+This script shows a minimal text-classification pipeline over the
+Hugging Face Banking77 dataset using the Synth LM interface. It also
+demonstrates a simple pre-send prompt-injection step as outlined in
+`synth_ai/learning/prompts/injection_plan.txt`.
+
+Notes
+- Network access is required to download the dataset and call the model.
+- Defaults to Groq with model `openai/gpt-oss-20b`.
+  - Export your key: `export GROQ_API_KEY=...`
+  - Override if needed: `export MODEL=openai/gpt-oss-20b VENDOR=groq`
+
+Run
+- `python -m synth_ai.learning.prompts.hello_world_in_context_injection_ex`
+
+What "in-context injection" means here
+- The script applies ordered substring replacements to the outgoing
+  `messages` array before calling the model. This mirrors the algorithm
+  described in `injection_plan.txt` without importing any non-existent
+  helper yet. You can adapt `INJECTION_RULES` to your needs.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import random
+
+from datasets import load_dataset
+
+# Use the v3 LM class present in this repo
+from synth_ai.lm.core.main_v3 import LM, build_messages
+
+# Use Overrides context to demonstrate matching by content
+from synth_ai.lm.overrides import LMOverridesContext
+from synth_ai.tracing_v3.abstractions import LMCAISEvent
+from synth_ai.tracing_v3.session_tracer import SessionTracer
+
+INJECTION_RULES = [
+    {"find": "accnt", "replace": "account"},
+    {"find": "atm", "replace": "ATM"},
+    {"find": "txn", "replace": "transaction"},
+]
+
+
+async def classify_sample(lm: LM, text: str, label_names: list[str]) -> str:
+    """Classify one Banking77 utterance and return the predicted label name."""
+    labels_joined = ", ".join(label_names)
+    system_message = (
+        "You are an intent classifier for the Banking77 dataset. "
+        "Given a customer message, respond with exactly one label from the list. "
+        "Return only the label text with no extra words.\n\n"
+        f"Valid labels: {labels_joined}"
+    )
+    user_message = f"Message: {text}\nLabel:"
+
+    # Build canonical messages; injection will be applied inside the vendor via context
+    messages = build_messages(system_message, user_message, images_bytes=None, model_name=lm.model)
+    resp = await lm.respond_async(messages=messages)
+    raw = (resp.raw_response or "").strip()
+    return raw
+
+
+async def main() -> None:
+    # Configurable model/provider via env, with sensible defaults
+    # Default to Groq hosting `openai/gpt-oss-20b`
+    model = os.getenv("MODEL", "openai/gpt-oss-20b")
+    vendor = os.getenv("VENDOR", "groq")
+
+    # Construct LM
+    lm = LM(model=model, vendor=vendor, temperature=0.0)
+
+    # Load Banking77 dataset
+    # Columns: {"text": str, "label": int}; label names at ds.features["label"].names
+    print("Loading Banking77 dataset (split='test')...")
+    ds = load_dataset("banking77", split="test")
+    label_names: list[str] = ds.features["label"].names  # type: ignore
+
+    # Sample a few items for a quick demo
+    n = int(os.getenv("N_SAMPLES", "8"))
+    idxs = random.sample(range(len(ds)), k=min(n, len(ds)))
+
+    correct = 0
+    # Apply overrides for all calls in this block (match by content)
+    overrides = [
+        {"match": {"contains": "atm", "role": "user"}, "injection_rules": INJECTION_RULES},
+        {"match": {"contains": "refund"}, "params": {"temperature": 0.0}},
+    ]
+    with LMOverridesContext(overrides):
+        for i, idx in enumerate(idxs, start=1):
+            text: str = ds[idx]["text"]  # type: ignore
+            gold_label_idx: int = int(ds[idx]["label"])  # type: ignore
+            gold_label = label_names[gold_label_idx]
+
+            try:
+                pred = await classify_sample(lm, text, label_names)
+            except Exception as e:
+                print(f"[{i}] Error calling model: {e}")
+                break
+
+            # Normalize and check exact match; if not exact, attempt a loose fallback
+            norm_pred = pred.strip().lower()
+            label_lookup = {ln.lower(): ln for ln in label_names}
+            pred_label = label_lookup.get(norm_pred)
+            if pred_label is None:
+                # Fallback: pick the label with highest substring overlap (very naive)
+                # This avoids extra deps; feel free to replace with a better matcher.
+                def score(cand: str) -> int:
+                    c = cand.lower()
+                    return sum(1 for w in c.split() if w in norm_pred)
+
+                pred_label = max(label_names, key=score)
+
+            is_correct = pred_label == gold_label
+            correct += int(is_correct)
+            print(
+                f"[{i}] text={text!r}\n    gold={gold_label}\n    pred={pred} -> mapped={pred_label} {'✅' if is_correct else '❌'}"
+            )
+
+    if idxs:
+        acc = correct / len(idxs)
+        print(f"\nSamples: {len(idxs)} | Correct: {correct} | Accuracy: {acc:.2%}")
+
+    # ------------------------------
+    # Integration tests (three paths)
+    # ------------------------------
+    print("\nRunning integration tests with in-context injection...")
+    test_text = "I used the atm to withdraw cash."
+
+    # 1) LM path with v3 tracing: verify substitution in traced messages
+    tracer = SessionTracer()
+    await tracer.start_session(metadata={"test": "lm_injection"})
+    await tracer.start_timestep(step_id="lm_test")
+    # Use a tracer-bound LM instance
+    lm_traced = LM(model=model, vendor=vendor, temperature=0.0, session_tracer=tracer)
+    with LMOverridesContext([{"match": {"contains": "atm"}, "injection_rules": INJECTION_RULES}]):
+        _ = await classify_sample(lm_traced, test_text, label_names)
+    # inspect trace
+    events = [
+        e
+        for e in (tracer.current_session.event_history if tracer.current_session else [])
+        if isinstance(e, LMCAISEvent)
+    ]
+    assert events, "No LMCAISEvent recorded by SessionTracer"
+    cr = events[-1].call_records[0]
+    traced_user = ""
+    for m in cr.input_messages:
+        if m.role == "user":
+            for part in m.parts:
+                if getattr(part, "type", None) == "text":
+                    traced_user += part.text or ""
+    assert "ATM" in traced_user, f"Expected substitution in traced prompt; got: {traced_user!r}"
+    print("LM path trace verified: substitution present in traced prompt.")
+    await tracer.end_timestep()
+    await tracer.end_session()
+
+    # 2) OpenAI wrapper path (AsyncOpenAI to Groq): ensure apply_injection is active
+    try:
+        import synth_ai.lm.provider_support.openai as _synth_openai_patch  # noqa: F401
+        from openai import AsyncOpenAI
+
+        base_url = os.getenv("OPENAI_BASE_URL", "https://api.groq.com/openai/v1")
+        api_key = os.getenv("OPENAI_API_KEY") or os.getenv("GROQ_API_KEY") or ""
+        client = AsyncOpenAI(base_url=base_url, api_key=api_key)
+        messages = [
+            {"role": "system", "content": "Echo user label."},
+            {"role": "user", "content": f"Please classify: {test_text}"},
+        ]
+        with LMOverridesContext(
+            [{"match": {"contains": "atm"}, "injection_rules": INJECTION_RULES}]
+        ):
+            _ = await client.chat.completions.create(
+                model=model, messages=messages, temperature=0
+            )
+        # Not all models echo input; instead, verify that our injected expectation matches
+        expected_user = messages[1]["content"].replace("atm", "ATM")
+        if messages[1]["content"] == expected_user:
+            print("OpenAI wrapper: input already normalized; skipping assertion.")
+        else:
+            print("OpenAI wrapper: sent message contains substitution expectation:", expected_user)
+    except Exception as e:
+        print("OpenAI wrapper test skipped due to error:", e)
+
+    # 3) Anthropic wrapper path (AsyncClient): ensure apply_injection is active
+    try:
+        import anthropic
+        import synth_ai.lm.provider_support.anthropic as _synth_anthropic_patch  # noqa: F401
+
+        a_model = os.getenv("ANTHROPIC_MODEL", "claude-3-5-haiku-20241022")
+        a_key = os.getenv("ANTHROPIC_API_KEY")
+        if a_key:
+            a_client = anthropic.AsyncClient(api_key=a_key)
+            with LMOverridesContext(
+                [{"match": {"contains": "atm"}, "injection_rules": INJECTION_RULES}]
+            ):
+                _ = await a_client.messages.create(
+                    model=a_model,
+                    system="Echo user label.",
+                    max_tokens=64,
+                    temperature=0,
+                    messages=[{"role": "user", "content": [{"type": "text", "text": test_text}]}],
+                )
+            print("Anthropic wrapper call completed (cannot reliably assert echo).")
+        else:
+            print("Anthropic wrapper test skipped: ANTHROPIC_API_KEY not set.")
+    except Exception as e:
+        print("Anthropic wrapper test skipped due to error:", e)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())