This is a PEFT LoRA adapter for unsloth/Qwen3-4B-Instruct-2507, trained for agentic code repair on the SWE-Gym moto split using a search/replace patch format and honest anchored retrieval.

This checkpoint is the best KL-GRPO continuation of the visible16k SFT adapter, selected at step 50 with beta 0.02 and learning rate 2e-6.

Checkpoint

Local source checkpoint:

/mnt/disks/unslothai/datta0/cache/qwen3-grpo-patch/20260604_115126_swegym_q4b-sftbest-kl02-lr2e6_787ca07/checkpoints/best_holdout

Training run:

20260604_115126_swegym_q4b-sftbest-kl02-lr2e6_787ca07

Training Summary

Base model: unsloth/Qwen3-4B-Instruct-2507
Initial adapter: visible16k SFT checkpoint
RL method: KL-GRPO continuation
Beta: 0.02
Learning rate: 2e-6
Max steps: 50
Best step: 50
Held-out greedy at selection: 8/35
Held-out mean reward at selection: 0.4169

Evaluation

Held-out SWE-Gym moto eval, honest anchored retrieval, search/replace patch format:

Run	greedy	pass@8	single pass@8	multi pass@8
original retriever sample 1	8/35	11/35	8/18	3/17
original retriever sample 2	8/35	12/35	8/18	4/17
original retriever sample 3	8/35	12/35	9/18	3/17
opsnake retriever sample	8/35	12/35	10/18	2/17

The default pass@8 decoding used for the positive samples was temperature=0.8, top_p=0.95. A higher-diversity probe at temperature=1.0, top_p=1.0 was negative: 10/35 overall pass@8 and 1/17 multi-file pass@8.

Intended Use

This is the strongest Qwen3-4B adapter artifact from the local investigation so far. It is intended for research on SWE-Gym style code repair with the same search/replace prompt contract and sandbox scoring pipeline.

Loading

Use this as a PEFT adapter on top of unsloth/Qwen3-4B-Instruct-2507.

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "unsloth/Qwen3-4B-Instruct-2507"
adapter = "imdatta0/qwen3-4b-swegym-moto-kl02-adapter"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

Limitations

Evaluated only on the local SWE-Gym moto held-out split used in this investigation.
Overall pass@8 is repeated at 11-12/35, but multi-file repair remains weak at 2-4/17 depending on retrieval/sample.
Metrics depend on the repository's retrieval, prompt, search/replace extraction, patch application, and sandbox scoring code.
This adapter requires the base model and is not a merged full model.