codellama-fine-tuning / QUICK_TEST_COMMAND.md
Prithvik-1's picture
Upload QUICK_TEST_COMMAND.md with huggingface_hub
597b65c verified
# πŸš€ Quick Test Command for Single Training Sample
## βœ… Updated Command (With Fixes Applied)
```bash
cd /workspace/ftt/codellama-migration && source /venv/main/bin/activate && python3 test_single_training_sample.py
```
**This will:**
- Load the first training sample
- Test with temperatures: 0.1, 0.2, 0.3
- Show expected vs generated output
- Help you find the best temperature
---
## πŸ”§ Manual Test Command (Custom Temperature)
```bash
cd /workspace/ftt/codellama-migration
source /venv/main/bin/activate
# Extract sample #1 instruction
INSTRUCTION=$(sed -n '1p' datasets/processed/split/train.jsonl | python3 -c "import sys, json; print(json.load(sys.stdin)['instruction'])")
# Test with optimized parameters
python3 scripts/inference/inference_codellama.py \
--mode local \
--model-path training-outputs/codellama-fifo-v1 \
--prompt "$INSTRUCTION" \
--max-new-tokens 1000 \
--temperature 0.1
```
---
## 🎯 Key Hyperparameter Changes Made
| Parameter | Old | New | Impact |
|-----------|-----|-----|--------|
| **Repetition Penalty** | 1.1 | **1.2** | βœ… Prevents repetitive text |
| **Prompt Format** | No EOS | **+ EOS token** | βœ… Matches training format |
| **Temperature** | 0.3 | **0.1 (recommended)** | βœ… More deterministic |
| **Response Extraction** | Full decode | **New tokens only** | βœ… Cleaner output |
---
## πŸ“Š Expected Results
With these changes, you should see:
- βœ… **Verilog code generation** (not text notes)
- βœ… **Complete modules** (module ... endmodule)
- βœ… **No repetitive "Note:" statements**
- βœ… **Better match with training data**
---
## πŸ” If Still Getting Text Instead of Code
Try these in order:
1. **Lower temperature to 0.05:**
```bash
--temperature 0.05
```
2. **Increase repetition penalty to 1.3:**
(Edit inference script or add as parameter)
3. **Check prompt format:**
- Ensure instruction matches training data exactly
- No extra formatting or wrappers
4. **Verify training data format:**
- Should be: `instruction + EOS + response + EOS`
- Response should start with ```verilog
---
**Test and let me know the results!**