e2h-polaris-qwen3-4b-full-mixed, step 50
GRPO checkpoint from the full-split round of the E2H (easy-to-hard) Polaris math runs, trained with a fork of verl 0.4.1 on Polaris-Dataset-53K problems split into easy and hard by Qwen/Qwen3-4B's own 16-sample pass rates. The training sets hold every problem in each band.
- Base model: Qwen/Qwen3-4B
- Reward: rule-based (correct against the reference answer) on the easy problems; on the hard ones, majority vote among each problem's 32 samples, judged with math_verify (the reference answer is never used)
- Training data:
train_all(train_easy+train_hard, 26,790 problems); 64 prompts per step ร 32 samples, lr 1e-6, KL loss 0.001 (low_var_kl), responses up to 8,192 tokens - This checkpoint: step 50. The run was set up for 1,200 steps and stopped after its first checkpoint, at step 50.
- Evaluated at steps 0 and 50 on
eval_all_small(850 problems, 50 at each of the base's 0-16 correct out of 16) andeval_hard_small(1,000 problems the base solved 0 times in 16), 8 samples each.
Load it with AutoModelForCausalLM.from_pretrained("davidafrica/e2h-polaris-qwen3-4b-full-mixed") or
vllm.LLM("davidafrica/e2h-polaris-qwen3-4b-full-mixed"); branch step-50 holds the same files. Weights are bf16,
merged from verl's FSDP shards.
- Downloads last month
- 23
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support