R_REDDYX.XYZ
LLMPython★ 1,243⑂ 163 forks

RikiyaT/lm_understand-logs

FULL LOGS OF LLM EXPERIMENTS UNDERSTANDING NEURAL NETWORK WEIGHTS

Complete logs of 8 experiments from the blog on LLMs understanding and writing neural network weights. Includes prompts, model responses, reasoning, code, and evaluations for tasks identifying games, generation, fixing, and creating networks from scratch.

// KEY FEATURES

  • 8 experiments: from game identification to writing a Go network
  • Full dialogues with gpt-6-astra: prompts, reasoning, code, outputs
  • Audit of 1125 turns: confirmed tool-free operation
  • Python sandbox for experiments 3, 6, 7, 8 with bubblewrap

// DEEP DIVE: HOW IT WORKS

The repository contains structured logs for each of the eight experiments described in the blog post. Every experiment folder includes a README mapping runs to the post's tables, plus prompt.txt, events.jsonl, answer.md, and tool_calls.jsonl. The gpt-6-astra model was invoked via Codex CLI with a set reasoning effort (high, xhigh, or max), while all built-in Codex tools were disabled via harness/codex_config.toml.

Inside are 1,125 dialogue turns, each audited for the absence of tool calls (the audit.tool_free field). For experiments 3, 6, 7, and 8 the model wrote Python code executed in an isolated sandbox harness/sandbox.py (bubblewrap, no network, with a timeout). Logs capture tokens, reasoning, code, and its output, enabling reproduction of the path from prompt to final weights.

Limitations: only runs reported in the post are included; trial runs of earlier prompt versions are omitted. Working with the logs does not require running models — this is a static dataset. Sandbox dependencies are described in harness/, but the repository itself contains no reproduction scripts, only the experiment artifacts.

// USE CASES

Best for: LLM researchers studying weight space programming and weight interpretability

  • Study how an LLM reasons about neural network weights without a forward pass
  • Reproduce prompt chains for weight space programming tasks
  • Analyze model error patterns when editing network weights
Open on GitHub →

New repositories every 30 minutes

REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.

Join on Telegram
← Full catalog·Full index

// SIMILAR REPOSITORIES

RISC-V CPU COMPILED INTO QWEN3.5 WEIGHTS★ 814STATIC CHAT FOR DEEPSEEK API★ 1701OPENLLM: RUN LLMS AS OPENAI API★ 12607OPEN-SOURCE FINANCIAL LLMS FOR DATA ANALYSIS★ 21591REA - REVERSE ENGINEERING WITHOUT BOUNDARIES★ 460891DeepSeek Plugin Framework★ 281341

← FULL CATALOG

RikiyaT/lm_understand-logs — Full Logs OF LLM Experiments…