Verification status
Every edited tensor passed both base and edited SHA256 checks against the upstream manifest. 72,986,982 BF16 values differ from the base. Short text generation passed on a local RTX 2080 SUPER using Transformers BF16. Independent benchmark accuracy is not yet available. This is not a claim of measured performance improvement.
The original multimodal architecture and unedited vision weights are retained; only the named text-model tensors were changed. Vision behavior has not been evaluated.
Provenance and licensing
Original model: https://huggingface.co/Qwen/Qwen3.5-0.8B (Apache-2.0).
Patch and reconstruction code: https://github.com/RikiyaT/lm_understand-logs/tree/4d09ba6d6a7e891dd8063974dfe81b3d78e3494f/exp8_math/runs/p10c/edit
Research by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab: https://rikiyat.github.io/blog/weight-space-programming/
Published patches/logs are CC BY 4.0; reconstruction scripts are MIT. Preserve the included upstream license and attribution. The number in p10c is the editor's forward-pass budget, not a training epoch.
Evaluation protocol
Planned local MATH-500 evaluation uses Transformers BF16, not the authors' vLLM engine. Thinking enabled; temperature 0.6, top_p 0.95, top_k 20, generated-token presence penalty 1.5, 32768 maximum new tokens, four responses per problem. Engine and RNG implementation differ from the paper. No local benchmark score is claimed before completion.