e2h-polaris-gemma-4-e4b-it-hard_mv, step 200
GRPO checkpoint from the E2H (easy-to-hard) Polaris math runs, trained with upstream verl 0.9.1 (v0 trainer, FSDP2) on Polaris-Dataset-53K problems split by google/gemma-4-E4B-it's own 16-sample pass rates (easy 10-15/16, hard 1-9/16).
- Base model: google/gemma-4-E4B-it, the full multimodal checkpoint; only the language model was trained, so the vision and audio towers are the base's
- Reward: majority vote: each of the 8 samples is rewarded for matching the most common final
\boxed{}answer among them - Training data:
train_hard; 200 steps of 64 prompts Γ 8 samples, lr 1e-6, KL loss 0.001 (low_var_kl), 2048-token responses with thinking on - This checkpoint: step 200 of 200
Each saved step (every 50) is a branch: load one with
AutoModelForImageTextToText.from_pretrained("davidafrica/e2h-polaris-gemma-4-e4b-it-hard_mv", revision="step-200")
or vllm.LLM("davidafrica/e2h-polaris-gemma-4-e4b-it-hard_mv", revision="step-200"). main holds the final step.
Weights are bf16, merged from verl's FSDP shards. Gemma 4 shares K/V across its
last 18 layers, so their own k_proj, v_proj and k_norm are never used; they are
copied unchanged from the base so the tensor names match it.
- Downloads last month
- 47
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support