Add model card for main-table Harness-R1 checkpoints
Browse files
README.md
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: transformers
|
| 4 |
+
base_model: Qwen/Qwen3.5-9B
|
| 5 |
+
tags:
|
| 6 |
+
- harness-r1
|
| 7 |
+
- agent
|
| 8 |
+
- reinforcement-learning
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# Harness-R1 Checkpoints
|
| 12 |
+
|
| 13 |
+
This repository hosts the two **Harness-R1 harness-engineer** checkpoints reported in the paper main table (`Ours` block):
|
| 14 |
+
|
| 15 |
+
| Subfolder | Paper row | Role |
|
| 16 |
+
|---|---|---|
|
| 17 |
+
| [`harness-r1-e0-iter143`](./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`. |
|
| 18 |
+
| [`harness-r1-e1-iter495`](./harness-r1-e1-iter495) | **Agent SFT + Harness-R1** | Continued engineer RL (E0→E1) on the Agent-SFT target (A1). Valid-selected checkpoint `iter495`. |
|
| 19 |
+
|
| 20 |
+
Both folders are full Hugging Face Transformers checkpoints (≈18GB each, single `model.safetensors`) derived from Qwen3.5-9B.
|
| 21 |
+
|
| 22 |
+
## Quick load
|
| 23 |
+
|
| 24 |
+
```python
|
| 25 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 26 |
+
|
| 27 |
+
repo = "ShaoShuai0605/Harness-R1"
|
| 28 |
+
# main-table Harness-R1 engineer
|
| 29 |
+
tok = AutoTokenizer.from_pretrained(repo, subfolder="harness-r1-e0-iter143", trust_remote_code=True)
|
| 30 |
+
model = AutoModelForCausalLM.from_pretrained(repo, subfolder="harness-r1-e0-iter143", trust_remote_code=True)
|
| 31 |
+
```
|
| 32 |
+
|
| 33 |
+
For the Agent-SFT + Harness-R1 engineer, use `subfolder="harness-r1-e1-iter495"`.
|
| 34 |
+
|
| 35 |
+
Code and evaluation harness: see the companion open-source release for Harness-R1.
|
| 36 |
+
|
| 37 |
+
## Notes
|
| 38 |
+
|
| 39 |
+
- These checkpoints are **harness engineers** (patch generators), not the frozen target agents used at evaluation time.
|
| 40 |
+
- Serving for tool-call / patch generation should follow the project’s Qwen3.5 engineer protocol (`enable_thinking` as used in the corresponding eval configs).
|