Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# LLaMA-3.1-8B SFT (No Prompt Masking)
|
| 2 |
+
|
| 3 |
+
Fine-tuned LLaMA-3.1-8B using SFT instruction tuning **without prompt masking** (loss computed on all tokens).
|
| 4 |
+
|
| 5 |
+
## Training Details
|
| 6 |
+
- **Base Model**: meta-llama/Llama-3.1-8B
|
| 7 |
+
- **Dataset**: UltraChat-200K + SafetyLlama (~200K examples)
|
| 8 |
+
- **Training**: 1 epoch (6726 steps)
|
| 9 |
+
- **Prompt Masking**: Disabled (loss on all tokens)
|
| 10 |
+
|
| 11 |
+
## Evaluation Results
|
| 12 |
+
|
| 13 |
+
| Benchmark | Baseline | This Model |
|
| 14 |
+
|-----------|----------|------------|
|
| 15 |
+
| GSM8K | 16.4% | 29.0% |
|
| 16 |
+
| MMLU | 58.1% | 58.4% |
|
| 17 |
+
| SST Safety | 62.0% | 78.0% |
|
| 18 |
+
| AlpacaEval | 1.57% | **5.3%** |
|
| 19 |
+
|
| 20 |
+
## Files
|
| 21 |
+
- `eval_baseline/`: Baseline evaluation results (pre-finetuning Llama-3.1-8B)
|
| 22 |
+
|
| 23 |
+
## Reference
|
| 24 |
+
Part of CS336 Assignment 5 (SFT Instruction Tuning). See [building-from-scratch/sft](https://github.com/garg-aayush/building-from-scratch/tree/main/sft) for details.
|