CoolFace
Modelpublic

imdatta0/qwen3-4b-swegym-moto-kl02-sft20k-hardmulti-interp-teachergap-v1-alpha025-adapter

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face
0likes5downloads
Model Card

Qwen3-4B SWE-Gym Moto Hardmulti Teachergap Interpolation Adapter

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

The adapter was produced as an arithmetic interpolation between the Qwen3-4B hard-multi 20k adapter and the teacher-gap-v1 adapter:

text
hardmulti + 0.25 * (teacher_gap_v1 - hardmulti)

It was evaluated on the held-out SWE-Gym Moto search/replace patch task with 20k anchored retrieval context.

Evaluation

Held-out sample seed 9012, k=1 smoke:

adaptercontextseedgreedyselected@1any-of-1single anymulti any
hard-multi/teacher-gap interp alpha 0.2520k90128/3510/3511/359/182/17

This was a negative/non-frontier checkpoint. It loaded and evaluated cleanly, but it did not recover the teacher-gap-v1 multi-file tail hit and it underperformed the hard-multi 20k frontier.

Contents

  • —adapter_model.safetensors: PEFT LoRA adapter weights
  • —adapter_config.json: PEFT adapter configuration
  • —tokenizer_config.json: tokenizer configuration
  • —interp_meta.json: local interpolation metadata

The base model weights are not included.