# LogSage Training Report

## Run

- Run directory: `runs/full-20260509-145106`
- Output directory: `LogSage-Qwen2.5-7B-QLoRA-v0`
- Base model: `unsloth/Qwen2.5-7B-Instruct-bnb-4bit`
- CUDA available: `True`
- GPU: `NVIDIA A10G`

## Dataset

- Rows: 1116
- Severity distribution: {'high': 348, 'low': 252, 'medium': 516}
- Max input characters: 2853

## Hyperparameters

- Epochs: 3.0
- Max sequence length: 2048
- Train batch size: 2
- Gradient accumulation steps: 4
- Learning rate: 0.0002
- Warmup steps: 10
- Eval steps: 25
- Save steps: 25

## Final Metrics

```json
{
  "train_runtime": 2058.2405,
  "train_samples_per_second": 1.463,
  "train_steps_per_second": 0.184,
  "total_flos": 5.54606820531671e+16,
  "train_loss": 0.7878258203072522,
  "epoch": 3.0,
  "eval_loss": 0.8106855750083923,
  "eval_runtime": 22.2714,
  "eval_samples_per_second": 5.029,
  "eval_steps_per_second": 1.257
}
```

## Observability Artifacts

- Console/file log: `runs/full-20260509-145106/train.log`
- TensorBoard: `runs/full-20260509-145106/tensorboard`
- Metrics JSONL: `runs/full-20260509-145106/metrics/metrics.jsonl`
- Sample eval generations: `runs/full-20260509-145106/eval_outputs.jsonl`

## Notes

This is a learning-grade fine-tune. Validate outputs manually before using them for incident decisions.