garg-aayush/llama31-8b-sft-mask

LLaMA-3.1-8B SFT (Prompt Masking)

Fine-tuned LLaMA-3.1-8B using SFT instruction tuning with prompt masking (loss computed only on response tokens).

Training Details

• Base Model: meta-llama/Llama-3.1-8B
• Dataset: UltraChat-200K + SafetyLlama (~200K examples)
• Training: 1 epoch (6326 steps)
• Prompt Masking: Enabled (loss on response tokens only)

Evaluation Results

Benchmark	Baseline	This Model
GSM8K	16.4%	32.7%
MMLU	58.1%	58.2%
SST Safety	62.0%	77.0%
AlpacaEval	1.57%	4.5%

Files

• eval_baseline/: Baseline evaluation results (pre-finetuning Llama-3.1-8B)

Reference

Part of CS336 Assignment 5 (SFT Instruction Tuning). See building-from-scratch/sft for details.