burnerqmatrixacl/qmatrix-codesft-adapters

Q-Matrix Code-SFT LoRA adapters

LoRA adapters (r=8, alpha=16) trained for the paper "Item-Level Coreset Selectors Choose a Corpus They Never Scored: A Q-Matrix Audit for Code Supervised Fine-Tuning." Each adapter is a separate selector x budget x seed cell, trained on a 5K/10K coreset of the 98,672-item Evol-Instruct + KodCode pool with the Magicoder @@ Instruction / @@ Response format, 2 epochs unless the name says otherwise.

Code, coresets, the Q-matrix, and the analysis that recomputes the paper's tables are in the companion dataset repository: https://huggingface.co/datasets/burnerqmatrixacl/qmatrix-codesft-artifacts

Naming

coder3b_*   base = Qwen/Qwen2.5-Coder-3B   (paper: Qwen2.5-Coder-3B-Base)
base1.5b_*  base = Qwen/Qwen2.5-1.5B       (paper: Qwen2.5-1.5B-Base)

<selector>_k<budget>_seed<42|43|44>[_ep<N>]
  cherry        Cherry-LLM IFD
  ifd_only      IFD-only ablation
  maxcov        MaxCov submodular coverage (epoch-curve: ep2/ep3/ep4)
  random        uniform random over the pool
  strat_rand    StratRand at the diagnosed mixture
  strat_alphaNN StratRand with evol-fraction alpha = 0.NN

alphaNN is the Evol fraction, so the KodCode fraction the paper calls alpha is 1 - NN/100. strat_alpha79 is the natural pool ratio, and strat_alpha50 sits near the diagnosed mixture.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-Coder-3B"            # or Qwen/Qwen2.5-1.5B for base1.5b_*
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16")
model = PeftModel.from_pretrained(model, "burnerqmatrixacl/qmatrix-codesft-adapters",
                                  subfolder="coder3b_cherry_k5000_seed42")
tok = AutoTokenizer.from_pretrained(base)

Adapters inherit the licenses of their base models and training data (Qwen2.5 / Evol-Instruct / KodCode); consult those sources for terms.