RikiyaT/Qwen3.5-0.8B-weight-space-programming
AI model · https://huggingface.co/RikiyaT/Qwen3.5-0.8B-weight-space-programming
First seen by Oriel on .
AI summary
This Qwen3.5-0.8B checkpoint shows that an LLM can improve a model’s math performance by directly editing a small subset of weights, without training or gradient descent.
- Astra used up to 50 practice-problem forward passes to identify a looping direction and subtract it from 64 output columns in each of layers 15 and 19.
- The edit changed 131,072 weights, or 0.016% of the model, and reduced the rate of generations hitting the token limit.
- Reported scores rose from 7.9% to 48.1% on MATH-500 and from 13.8% to 67.1% on GSM8K; IFEval rose from 21.0% to 62.3%.
- An independent 2,000-generation MATH-500 run scored 51.0%, with 31.6% of answers hitting the token limit; the base model scored 5.0% in a 25-problem smoke test.
- The published bit-exact diff and rebuild code verify that only the two specified down_proj tensors differ from the base model; the checkpoint is a full multimodal release.
#large language models #weight editing #math benchmarks #qwen
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Likes
Total likes reported by the source.
- Hugging Face Models: 132, latest observation ; +32 over observed period since
Downloads (30d)
Rolling 30-day downloads. A change in this value is not new downloads during the observed period.
- Hugging Face Models: 3,865, latest observation ; +1,897 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Models | RikiyaT/Qwen3.5-0.8B-weight-space-programming | https://huggingface.co/RikiyaT/Qwen3.5-0.8B-weight-space-programming Read discussion |