You can deploy this model and start earning money today!

shengjia-toronto/
r1distill-qwen1.5b-drgrpo-noent-cl20-step231-run10

Name: r1distill-qwen1.5b-drgrpo-noent-cl20-step231-run10
Availability: InStock

safetensors

qwen2

Views

No views yet

r1distill-qwen1.5b-drgrpo-noent-cl20-step231-run10

Dr.GRPO RL fine-tune of deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B on math, exported from a single training step for evaluation purposes.

Source

Local checkpoint path (FSDP-sharded, 8 ranks): /opt/dlami/nvme/workspace/shapley_workspace/shapley/checkpoints/shapley_baselines/drgrpo_noent_cl20_r1distill_qwen1.5b_math_ec2_run10_0430_1638/global_step_231
Step: 231
Run tag: run10_0430_1638
Training script: train_drgrpo_noent_cl20_r1distill_qwen1.5b_math_ec2.sh
Merge: python -m verl.model_merger merge --backend fsdp --local_dir <ckpt>/actor

Training

Base model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
Algorithm: Dr.GRPO (no per-group std normalization)
Entropy bonus: 0 (noent)
Clip ratio: 0.2 / 0.2 (cl20)
Rollouts per prompt: 8 (train), 256 (eval pass@k)
Train batch size: 256
Train sampling: temperature=1.0, top_p=1.0
Training data: DeepScaleR-Preview-Dataset (39,207 problems)

Project

Part of the Shapley@k RLVR research project (shengjia-toronto/shapley-experiments). This is a baseline checkpoint (Dr.GRPO), used for comparison against Shapley-shaped credit assignment.

License

Inherits from base model (deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B).

shengjia-toronto/r1distill-qwen1.5b-drgrpo-noent-cl20-step231-run10