Search-R1-Qwen3-8B-SDAR

A Qwen/Qwen3-8B policy trained with GRPO on Search-R1 multi-turn retrieval QA, with SDAR self-distillation.

The repository root holds the final policy, at training step 400.

The policy was trained for 400 steps with the Search-R1 tool-call interaction format. The raw FSDP checkpoint at step 400 is included under fsdp/ for resuming training.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("wckwan/Search-R1-Qwen3-8B-SDAR")
tokenizer = AutoTokenizer.from_pretrained("wckwan/Search-R1-Qwen3-8B-SDAR")

# An intermediate checkpoint
model_step = AutoModelForCausalLM.from_pretrained("wckwan/Search-R1-Qwen3-8B-SDAR", subfolder="step_20")

Checkpoints

• step_20/
• step_40/
• step_60/
• step_80/
• step_100/
• step_120/
• step_140/
• step_160/
• step_180/
• step_200/
• step_220/
• step_240/
• step_260/
• step_280/
• step_300/
• step_320/
• step_340/
• step_360/
• step_380/

Raw FSDP checkpoints

fsdp/global_step_<N>/ holds the unmerged verl FSDP checkpoint (sharded fp32 model state, optimizer state and extra state) for step(s) 400. These are for resuming training, not for inference — use the merged exports above to load a policy.