Harness-R1 Checkpoints

This repository hosts the two Harness-R1 harness-engineer checkpoints reported in the paper main table (Ours block):

Subfolder	Paper row	Role
harness-r1-e0-iter143	Harness-R1	Engineer trained with identity-clean SFT + online RL on frozen vanilla Qwen3.5-9B (A0). Valid-selected checkpoint iter143.
harness-r1-e1-iter495	Agent SFT + Harness-R1	Continued engineer RL (E0→E1) on the Agent-SFT target (A1). Valid-selected checkpoint iter495.

Both folders are full Hugging Face Transformers checkpoints (≈18GB each, single model.safetensors) derived from Qwen3.5-9B.

Quick load

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "ShaoShuai0605/Harness-R1"
# main-table Harness-R1 engineer
tok = AutoTokenizer.from_pretrained(repo, subfolder="harness-r1-e0-iter143", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, subfolder="harness-r1-e0-iter143", trust_remote_code=True)

For the Agent-SFT + Harness-R1 engineer, use subfolder="harness-r1-e1-iter495".

Code and evaluation harness: see the companion open-source release for Harness-R1.

Notes

• These checkpoints are harness engineers (patch generators), not the frozen target agents used at evaluation time.
• Serving for tool-call / patch generation should follow the project’s Qwen3.5 engineer protocol (enable_thinking as used in the corresponding eval configs).