Downloads · 30 days
0
logan7000/mllm-open-r1-gt-internvl35-8b-mmupt-full
mllm-open-r1-gt-internvl35-8b-mmupt-full is a machine learning model from logan7000. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Recipe: mmupt (beta 0.01, K 10, T 0.7, maxcompletionlength 2048, lr 1e-6, cosinewithminlr 0.1, warmupratio 0, weightdecay 0.01, maxgradnorm 1.0, bnpo, scalerewards group, effective batch 120 = 12 prompts/step, 1 epoch…
Downloads · 30 days
0
Access
Public
Updated Sep 11, 2026
Repo size
34.1 GB
Likes
0
Public
Click a slice to open those files.
.safetensors34.1 GB · 100%
From the Hugging Face model README
Recipe: mmupt (beta 0.01, K 10, T 0.7, max_completion_length 2048, lr 1e-6, cosine_with_min_lr 0.1, warmup_ratio 0, weight_decay 0.01, max_grad_norm 1.0, bnpo, scale_rewards group, effective batch 120 = 12 prompts/step, 1 epoch = 640 steps), identical to the big-tier reference script examples/mmr1_qwen25vl7b_gt_mmupt.sh (InternVL column). Trained on 4 GPUs (BS 3 x GA 10 x 4 = 120). best/ = best-by-val (MathVista-150 holdout) step 480; endpoint/ = checkpoint-640 (end of 1 epoch); training/ = train.log, trainer_state_best, trainer_state_endpoint, best_metric. Job 315846 on the JHU a100 cluster, 2026-09-07 00:53 -> 09-08 00:56 (24h03m). Eval protocol for the paper tables: v2 (T=0, top_p 0.95, 16k tokens, boxed prompt, mathruler rule + Qwen2.5-32B judge), endpoint selection, AVG4 over MathVista/MathVision/MathVerse/We-Math. Local source: /weka/scratch/jhu/dssg2026-ext-rghani1/yyang331/mllm-repro-out/mllm-co-grpo-dp/openr1_internvl35_8b_gt_mmupt Trained 2026-09-03..09-05 on the JHU a100 cluster.