imdatta0/qwen3-4b-swegym-moto-hardmulti-sft20k-teachergap-v1-adapter

Qwen3-4B SWE-Gym Moto Hardmulti Teacher-Gap V1 Adapter

This repository contains a PEFT LoRA adapter for unsloth/Qwen3-4B-Instruct-2507.

The adapter was trained as a bounded SFT continuation from the hard-multi 20k Qwen3-4B frontier adapter. The training mix preserved all 18 hard-multi rows and added 62 train-only teacher-gap rows selected from existing Coder-30B passing labels.

Evaluation

Held-out SWE-Gym Moto search/replace patch evaluation, 20k anchored retrieval context, bfloat16, sample seed 9012:

adapter	context	seed	greedy	selected@1	pass@8	single pass@8	multi pass@8
hard-multi plus teacher-gap SFT	20k	9012	10/35	11/35	14/35	9/18	5/17

This was the first measured 4B checkpoint in this investigation with 5/17 multi-file pass@8 in a single seed, gaining moto-6641 versus the seed9012 hard-multi frontier. It is not promoted over the hard-multi frontier overall because overall pass@8 drops from 16/35 to 14/35.

Contents

• adapter_model.safetensors: PEFT LoRA adapter weights
• adapter_config.json: PEFT adapter configuration
• checkpoint_metadata.json: local training metadata
• tokenizer files and chat template copied with the checkpoint

The base model weights are not included.