CoolFace
Modelpublic

xsample/Llama-3.1-MIG-Tulu-3-8B-SFT

sourceHugging Faceapache-2.0updated 1y agoView on Hugging Face
3likes9downloads
Model Card

Llama-3.1-MIG-Tulu-3-8B-SFT

Project | Github | Paper | HuggingFace's collection

Llama-3.1-MIG-Tulu-3-8B-SFT is fine-tuned on automatically selected 50K data.

Performance

MethodData SizeARCBBHGSMHEMMLUIFEvalAvg_objAEMTWildAvg_subAvg
Pool939K69.1563.8883.4063.4165.7767.1068.798.946.86-24.6638.4053.59
Random50K74.2464.8070.3651.2263.8661.0064.258.57<u>7.06</u>-22.1539.3651.81
ZIP50K77.6363.0052.5435.9865.0061.0059.196.716.64-32.1035.6947.44
IFD50K75.9363.5661.0349.3964.3953.6061.3212.307.03-20.2040.8351.08
#InsTag50K72.5464.8069.8348.1763.5065.9964.146.586.84-20.7038.2151.17
DEITA50K78.9866.1174.0749.3964.0064.33<u>66.15</u>10.196.83<u>-19.95</u>39.5052.83
CaR50K78.9869.0471.4252.4465.1556.7565.6312.556.95-20.6740.5753.10
QDIT50K<u>79.66</u>65.4270.74<u>53.05</u><u>65.06</u>57.3065.2115.786.76-20.56<u>41.03</u><u>53.12</u>
MIG50K80.00<u>66.39</u><u>72.02</u>57.9364.44<u>65.06</u>67.64<u>14.66</u>7.32-17.7742.9955.32

Citation

@article{chen2025mig,
  title={MIG: Automatic Data Selection for Instruction Tuning by Maximizing Information Gain in Semantic Space},
  author={Chen, Yicheng and Li, Yining and Hu, Kai and Ma, Zerun and Ye, Haochen and Chen, Kai},
  journal={arXiv preprint arXiv:2504.13835},
  year={2025}
}