self-play-pretrain-zero-data 0.0.1__py3-none-any.whl

self_play_pretrain_zero_data/__init__.py

@@ -0,0 +1,6 @@
+from self_play_pretrain_zero_data.self_play import (
+    SelfPlay,
+    ProgramExecutor
+)
+
+from self_play_pretrain_zero_data.brainfuck import Brainfuck

self_play_pretrain_zero_data/brainfuck.py

@@ -0,0 +1,96 @@
+from __future__ import annotations
+
+import random
+from itertools import chain, repeat
+
+from self_play_pretrain_zero_data.self_play import ProgramExecutor
+
+# single byte macro instructions from appendix E (table 4)
+
+MACROS = {
+    'Z': '[-]',
+    'R': '[->+<]',
+    'L': '[->+++<]',
+    'N': '[-<->]',
+    'C': '[->+>+<<]',
+    'G': '[>]',
+    'H': '[<]',
+    'W': '[[-]>+<]',
+    'V': '[.>]',
+    'X': '[-]++++++++++++++++',
+}
+
+
+def bracket_jumps(code):
+    jumps, stack = {}, []
+
+    for i, c in enumerate(code):
+        if c == '[':
+            stack.append(i)
+        elif c == ']' and stack:
+            left = stack.pop()
+            jumps[left] = i
+            jumps[i] = left
+
+    return jumps
+
+class Brainfuck(ProgramExecutor):
+    """appendix E execution semantics - every string is executable
+
+    - circular tape, cells modulo `modulus`
+    - unmatched brackets are no-ops
+    - `,` reads the input string, then uniform random bytes (random input tape omega)
+    - halts at `F`, the step budget, or the output cap
+    """
+
+    def __init__(
+        self,
+        tape_size = 1024,
+        modulus = 256,
+        max_steps = 100_000,
+        max_output_len = 1024
+    ):
+        self.tape_size = tape_size
+        self.modulus = modulus
+        self.max_steps = max_steps
+        self.max_output_len = max_output_len
+
+    def __call__(
+        self,
+        program: str,
+        input: str | bytes = '',
+        seed: int | None = None
+    ) -> str:
+        rng = random.Random(seed)
+
+        code = ''.join(MACROS.get(c, c) for c in program.partition('F')[0])
+        jumps = bracket_jumps(code)
+
+        input_bytes = input.encode('latin-1') if isinstance(input, str) else input
+
+        reads = chain(
+            (b % self.modulus for b in input_bytes),
+            (rng.randrange(self.modulus) for _ in repeat(None))
+        )
+
+        tape = [0] * self.tape_size
+        out = bytearray()
+        pc = ptr = 0
+
+        for _ in range(self.max_steps):
+            if pc >= len(code) or len(out) >= self.max_output_len:
+                break
+
+            match code[pc]:
+                case '>': ptr = (ptr + 1) % self.tape_size
+                case '<': ptr = (ptr - 1) % self.tape_size
+                case '+': tape[ptr] = (tape[ptr] + 1) % self.modulus
+                case '-': tape[ptr] = (tape[ptr] - 1) % self.modulus
+                case '.': out.append(tape[ptr])
+                case ',': tape[ptr] = next(reads)
+                case '[' if not tape[ptr]: pc = jumps.get(pc, pc)
+                case ']' if tape[ptr]: pc = jumps.get(pc, pc)
+
+            pc += 1
+
+        return out.decode('latin-1')

self_play_pretrain_zero_data/self_play.py

@@ -0,0 +1,42 @@
+from __future__ import annotations
+from typing import Iterable
+from abc import ABC, abstractmethod
+
+from torch.nn import Module
+
+# executor
+
+class ProgramExecutor(ABC):
+    """maps a program string to its output string"""
+
+    @abstractmethod
+    def __call__(self, program: str, **kwargs) -> str:
+        raise NotImplementedError
+
+# main class
+
+class SelfPlay(Module):
+    def __init__(
+        self,
+        generator: Module,
+        learner: Module,
+        executor: ProgramExecutor
+    ):
+        super().__init__()
+        self.generator = generator
+        self.learner = learner
+        self.executor = executor
+
+    def execute(
+        self,
+        programs: str | Iterable[str],
+        **kwargs
+    ) -> str | list[str]:
+
+        if isinstance(programs, str):
+            return self.executor(programs, **kwargs)
+
+        return [self.executor(program, **kwargs) for program in programs]
+
+    def forward(self, *args, **kwargs):
+        raise NotImplementedError

self_play_pretrain_zero_data-0.0.1.dist-info/METADATA

@@ -0,0 +1,77 @@
+Metadata-Version: 2.5
+Name: self-play-pretrain-zero-data
+Version: 0.0.1
+Summary: Implementation of Self-Play Pretraining with Zero Data
+Project-URL: Homepage, https://pypi.org/project/self-play-pretrain-zero-data/
+Project-URL: Repository, https://github.com/lucidrains/self-play-pretrain-zero-data
+Author-email: Phil Wang <lucidrains@gmail.com>
+License: MIT License
+        
+        Copyright (c) 2026 Phil Wang
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+License-File: LICENSE
+Keywords: artificial intelligence,deep learning,pretraining,self-play,synthetic data
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: einops>=0.8.1
+Requires-Dist: einx>=0.3.0
+Requires-Dist: torch-einops-utils>=0.0.29
+Requires-Dist: torch>=2.5
+Requires-Dist: x-transformers
+Provides-Extra: examples
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+## self-play-pretrain-zero-data (wip)
+
+Implementation of [Self-Play Pretraining with Zero Data](https://arxiv.org/abs/2609.30063)
+
+[Paper Review from @hu-po](https://www.youtube.com/watch?v=mGMiiPpWBSo)
+
+## Citations
+
+```bibtex
+@misc{cowsik2026selfplaypretrainingzerodata,
+    title    = {Self-Play Pretraining with Zero Data},
+    author   = {Aditya Cowsik and Kfir Dolev and Michael Y. Li and G. Bruno De Luca and Nourya Cohen and Noah D. Goodman and Yoav Levine},
+    year     = {2026},
+    eprint   = {2609.30063},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.AI},
+    url      = {https://arxiv.org/abs/2609.30063},
+}
+```
+
+```bibtex
+@misc{bloem2025universalpretrainingiteratedrandom,
+    title   = {Universal pre-training by iterated random computation},
+    author  = {Peter Bloem},
+    year    = {2025},
+    eprint  = {2506.20057},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/2506.20057},
+}
+```

self_play_pretrain_zero_data-0.0.1.dist-info/RECORD

@@ -0,0 +1,7 @@
+self_play_pretrain_zero_data/__init__.py,sha256=lT_TuvY_Z9sLWo62lXWI-ceLjCGaxOgCELddidon5cM,151
+self_play_pretrain_zero_data/brainfuck.py,sha256=edAtQOIpG5ggkb5J76eJx1O_HopVWvPp9AgYX7VyiwI,2627
+self_play_pretrain_zero_data/self_play.py,sha256=c1K2bwfoaxyvaHeqfJc_4899xaslM4F05KnJDZ8ib7A,978
+self_play_pretrain_zero_data-0.0.1.dist-info/METADATA,sha256=3Pq10qGXgwNBT4SlcEMMCxjowl2R_sul36clK-3rRuk,3209
+self_play_pretrain_zero_data-0.0.1.dist-info/WHEEL,sha256=W3fkpkm7-wf9vBI5Z-7s0eWkeM-spu78I8Neb98DeEg,87
+self_play_pretrain_zero_data-0.0.1.dist-info/licenses/LICENSE,sha256=e6AOF7Z8EFdK3IdcL0x0fLw4cY7Q0d0kNR0o0TmBewM,1066
+self_play_pretrain_zero_data-0.0.1.dist-info/RECORD,,

self_play_pretrain_zero_data-0.0.1.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.32.4
+Root-Is-Purelib: true
+Tag: py3-none-any

self_play_pretrain_zero_data-0.0.1.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Phil Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.