Spaces:

ML-Intern-lab
/

weight-space-programming-demo

Running on Zero

App Files Files Community

weight-space-programming-demo / README.md

akhaliq HF Staff

new model repo name + full MATH-500 numbers

edbdf0f verified 1 day ago

preview code

Raw History Blame Contribute Delete

1.2 kB

A newer version of the Gradio SDK is available: 6.31.0

Upgrade

metadata

title: Weight-Space-Programmed Qwen3.5-0.8B
emoji: 🧠
colorFrom: blue
colorTo: green
sdk: gradio
app_file: app.py
pinned: false
tags:
  - ml-intern
  - weight-space-programming

🧠 Weight-Space-Programmed Qwen3.5-0.8B

Live demo of the model from "Sparks of Weight Space Programming" — Qwen3.5-0.8B with 131,072 weights hand-edited by an LLM (no training) to remove its repetition-loop pathology. Type a math problem and watch the base model loop while the edited model solves it.

Model: akhaliq/Qwen3.5-0.8B-weight-space-programming (bit-exact, sha256-verified rebuild of the post's p50c edit)
Logs/diffs: RikiyaT/lm_understand-logs
Verified: full MATH-500 (500 × 4) — edited model 51.0% acc, 31.6% truncated (post: 48.1 / 33.1); base 5% / 95% on a 25-problem smoke (post: 7.9 / 93.0). Raw scores in eval/ of the model repo