Downloads · 30 days
0
logan7000/mllm-cogrpo-heter-internvl35-2b-x-gemma3-4b-mmr1-mmupt-groupA-internvl35-2b-full
mllm-cogrpo-heter-internvl35-2b-x-gemma3-4b-mmr1-mmupt-groupA-internvl35-2b-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 Gemma variant (beta 0.01, K 10, T 1.0, maxcompletionlength 1024, lr 1e-6, cosinewithminlr 0.1, warmupratio 0, weightdecay 0.01, maxgradnorm 1.0, bnpo, scalerewards group, effective batch 120 = 12 prompts…
Downloads · 30 days
0
Access
Public
Updated Sep 5, 2026
Repo size
9.4 GB
Likes
0
Public
Click a slice to open those files.
.safetensors9.4 GB · 100%
From the Hugging Face model README
Recipe: mmupt Gemma variant (beta 0.01, K 10, T 1.0, max_completion_length 1024, 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). Both sides of this co-learning pair use the same launcher, so the recipe is shared. best/ = best-by-val (MathVista-150 holdout) step 400; endpoint/ = checkpoint-481 (end of 1 epoch); training/ = train.log, trainer_state_best, trainer_state_endpoint, best_metric. Peer: Gemma-3-4B (group B repo of the same run). 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/mmr1_heter_internvl35_2b_x_gemma3_4b_mmuptv2_20260904_040127/model_a Trained 2026-09-03..09-05 on the JHU a100 cluster.