Downloads · 30 days
0
Rank1-Reconstruction-Qwen2.5-Math-1.5B
weizhepei/Rank1-Reconstruction-Qwen2.5-Math-1.5B
Rank1-Reconstruction-Qwen2.5-Math-1.5B is a machine learning model from weizhepei. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Oracle rank-1 reconstruction of the RLVR weight-delta trajectory: thetarecon = theta0 + v1 v1^T (thetastep - theta0), v1 = top right singular vector of the full 500-step observed delta trajectory (--mode reconstruct -…
Access
Public
Updated Jul 28, 2026
Repo size
3.1 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors3.1 GB · 100%
At a glance
- Access
- Public
- Created
- Jul 28, 2026
- Updated
- Jul 28, 2026
- SHA
- 21eba91b
Model card
From the Hugging Face model README
Rank-1 SVD Reconstruction — Qwen2.5-Math-1.5B (RELEX rebuttal artifact)
Oracle rank-1 reconstruction of the RLVR weight-delta trajectory:
theta_recon = theta_0 + v_1 v_1^T (theta_step - theta_0), v_1 = top right singular
vector of the full 500-step observed delta trajectory (--mode reconstruct --rank 1).
Base model = Qwen/Qwen2.5-Math-1.5B. Subfolders step100/ .. step500/.
Eval config (MUST match to reproduce)
TianHongZXY/MATH test, greedy temperature=0.0, top_p=0.95, top_k=-1, num_generation=1,
max_tokens=4096 (5000 problems). Accuracy = mean of the boolean label field.
Reference numbers to confirm
Rank-1 recon: 100=67.20 200=70.62 300=70.68 400=71.54 500=71.06 (RLVR real: 67.96/70.14/70.78/71.38/71.54 ; Rank-5 at weizhepei/Rank5-Reconstruction-Qwen2.5-Math-1.5B)
- Created
- Jul 28, 2026
- Updated
- Jul 28, 2026