austindixson/agent-dataset-hybrid

# Agent Training Dataset (Unsloth Format)

Hybrid dataset combining real Claude conversations with lightweight augmentations.

## Overview

- **Total conversations:** 4,747
- **Training examples:** 3,797 (train.jsonl)
- **Validation examples:** 475 (valid.jsonl)
- **Test examples:** 475 (test.jsonl)

## Format

ShareGPT format (compatible with Unsloth FastLanguageModel):

\`\`\`json
{
"conversations": [
{"from": "human", "value": "..."},
{"from": "gpt", "value": "..."}
]
}
\`\`\`

## Usage

### With Unsloth

\`\`\`python
from unsloth import FastLanguageModel
from datasets import load_dataset

# Load dataset
dataset = load_dataset(
"json",
data_files={
"train": "train.jsonl",
"valid": "valid.jsonl",
"test": "test.jsonl"
}
)

# Load model
model, tokenizer = FastLanguageModel.from_pretrained(
"unsloth/llama-3-8b-bnb-4bit",
load_in_4bit=True
)

# Train...
\`\`\`

### On PC with RTX 3060

\`\`\`bash
git lfs install
git lfs pull

python train.py \
--data train.jsonl \
--valid_data valid.jsonl
\`\`\`

## Data Composition

- **Real data:** 3,165 Claude Code conversations
- **Augmented:** 1,582 variations (reverse, shorten, paraphrase)
- **Total turns:** 150,619
- **Avg turns/conversation:** 31.7

## License

Same as source Claude Code conversations.