training/
├── configs/          # Training configurations
│   └── llama2-7b-lora.yaml
├── data/            # Prepared datasets (train/test splits)
├── output/          # Trained model outputs
├── scripts/         # Training and inference scripts
│   ├── prepare_dataset.py
│   ├── train.py
│   └── inference.py
└── notebooks/       # Jupyter notebooks for analysis

Prerequisites

Python 3.10+
GPU with 40GB+ VRAM (A100 recommended)
64GB+ system RAM
100GB+ free disk space

Installation

cd scripts/lora_training/training

# Create virtual environment
python -m venv venv
source venv/bin/activate

# Install dependencies
pip install transformers datasets trl peft accelerate bitsandbytes

# Or use conda
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
pip install transformers datasets trl peft accelerate bitsandbytes

Usage

1. Prepare Dataset

python scripts/prepare_dataset.py --input ../combined_20k.jsonl --output data

This splits the 20k dataset into train/test sets (95/5).

2. Train Model

python scripts/train.py --config configs/llama2-7b-lora.yaml

Or with custom parameters:

python scripts/train.py --config configs/llama2-7b-lora.yaml --epochs 5 --batch-size 8

Training takes approximately 6-24 hours depending on GPU.

3. Generate Summaries

python scripts/inference.py \
    --model output/llama2-7b-lora \
    --task "Fix parser crash on malformed JSON" \
    --files src/parser.cpp \
    --tests-run --test-count 294 \
    --commit --push

Configuration

Edit configs/llama2-7b-lora.yaml to adjust:

Base model: Change base_model to use different foundation models
LoRA rank: Adjust lora_r (higher = more capacity, slower)
Learning rate: Tune learning_rate
Epochs: Change num_train_epochs
Batch size: Adjust per_device_train_batch_size

Dataset Format

The dataset (combined_20k.jsonl) contains 20,000 examples with:

task: Task description
files_changed: List of changed files
tests_run: Boolean
commit: Boolean
push: Boolean
errors_seen: Boolean
test_count: Number of tests
output: Expected Cyron summary

Output

Trained model is saved to output/llama2-7b-lora/:

adapter_model.safetensors - LoRA weights
config.json - Model configuration
tokenizer.json - Tokenizer
peft_config.json - LoRA configuration

Troubleshooting

Out of memory:

Reduce per_device_train_batch_size
Enable gradient_checkpointing: true
Use load_in_4bit: true (already enabled)

Training loss not decreasing:

Lower learning rate (try 1e-4)
Increase warmup ratio
Check dataset quality

Generation too long/short:

Adjust max_new_tokens in inference script
Tune temperature and top_p

License

This training infrastructure is part of the AgenX project.