README.md

metadata

title: Weight-Space Programming Lab
emoji: 🧮
colorFrom: gray
colorTo: indigo
sdk: static
app_file: index.html
pinned: false
short_description: 'Hands-on labs: fix a small LLM by editing its weights.'

Weight-space programming lab

This Space holds nine hands-on labs that reproduce and probe Experiment 8 of "Sparks of Weight-Space Programming". In that experiment, an AI agent fixed the looping of Qwen3.5-0.8B. It edited 131,072 weights with arithmetic and did no training.

• Read the labs on the live site.
• Start with the Quick start page.
• Browse the code in the code/ folder.

Get the code

The first line below installs uv, which runs the lab's Python scripts with the packages they need and can fetch Python itself. The second line puts uv on your PATH (the folders the shell searches for commands) in this terminal. The uvx command then runs hf, the download tool from huggingface_hub, without installing it, and fetches the code/ folder:

curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env
uvx --from huggingface_hub hf download thegovind/weight-space-programming \
  --repo-type space --include "code/*" --local-dir ~/weight-space-programming
cd ~/weight-space-programming/code

If you already have uv, skip the first two lines.

Credits and licenses

• The project reproduces Experiment 8 of the blog post "Sparks of Weight-Space Programming" by the authors of "LLMs can understand and write neural network weights". It uses their public logs.
• The agent's 8 texts in code/labs/data/astra_pairs.json and the quoted log excerpts come from RikiyaT/lm_understand-logs. They are copyright the authors of "LLMs can understand and write neural network weights" and licensed CC BY 4.0. We extracted them from the logs. The selection and the JSON layout are ours. The labs, measurements and figures come from our own runs.
• code/wsp/grading.py includes answer-extraction and grading code copied from the paper's harness (exp8_math/code/mathqa.py). It is used under the MIT License. See code/NOTICE.md.
• Benchmarks: MATH-500 (from MATH, MIT License), GSM8K (MIT License) and IFEval (Apache License 2.0). The benchmark files are not included. code/scripts/prepare_data.py downloads them again from their public copies on Hugging Face. See code/data/README.md.
• No license is granted yet for this project's own code and text.