Add cold-start SFT metadata and inference config

615b9c3 verified about 1 month ago

1.19 kB

library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen3.5-9B
tags:
  - harness-engineering
  - agent
  - supervised-fine-tuning
  - reinforcement-learning

Harness-R1

Model artifacts for Harness-R1 experiments. The repository separates the target agent from the Harness Engineer so that the two roles are not confused.

Directory	Role
`cold-start-sft/`	Supervised-only Harness Engineer (SFT/0), used to initialize E0 RL
`harness-r1/`	E0 Harness Engineer for the vanilla A0 target agent
`agent-sft-harness-r1/`	E1 Harness Engineer conditioned on A1 trajectories
`target-agent-sft/`	Direct Agent-SFT target model (A1), not a Harness Engineer

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "ShaoShuai0605/Harness-R1"
subfolder = "harness-r1"  # Or cold-start-sft / agent-sft-harness-r1 / target-agent-sft

tokenizer = AutoTokenizer.from_pretrained(
    repo,
    subfolder=subfolder,
    trust_remote_code=True,
)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    subfolder=subfolder,
    trust_remote_code=True,
)

See the README inside each directory for its intended role.