Terminal-Lego: What Makes Interaction Trajectories Effective for Training Terminal Agents?

🤗 Instances • 🤗 Trajectories • 🤗 Terminal-Lego-Qwen3-8B/32B • 🧑‍💻 Code • 📖 Paper

📖 Project

1. 🔭 Pipeline Overview

2. 🚀 Quick Start

2.1 Install dependencies

pip install -r requirements.txt

2.2 Configure

cp .env.example .env
# Edit .env with your API keys

Required environment variables:

• OPENAI_API_KEY — API key for an OpenAI-compatible LLM service
• OPENAI_API_BASE — API base URL (default: https://api.openai.com/v1)
• SO_API_KEY — StackOverflow API key (get one here)
• MODEL_NAME — Model to use (default: claude-opus-4-6)

2.3 Run the full pipeline

bash run_pipeline.sh

This runs in a loop: scrape → generate → validate → next round.

3. 🔧 Running Steps Individually

3.1 Scrape StackOverflow

python scraper/so_scraper.py \
    --round 1 \
    --output ./data \
    --count 5000 \
    --api-key "$SO_API_KEY"

3.2 Generate Tasks

python generator/task_generator.py \
    --input ./data/so_data_r1.json \
    --output ./data/candidates_r1 \
    --workers 16 \
    --model claude-opus-4-6

3.3 Validate Tasks (requires Docker)

python validator/validate_tasks.py \
    --input ./data/candidates_r1 \
    --output ./data/validated_r1 \
    --workers 8 \
    --timeout 300

4. 📦 Output Format

Each generated task has this structure:

task_00001/
├── instruction.md          # Task description
├── task.toml               # Metadata (difficulty, category, tags)
├── environment/
│   ├── Dockerfile          # Container setup
│   └── task_file/          # Input files
├── solution/
│   └── solve.sh            # Reference solution
└── tests/
    ├── test.sh             # Test runner
    └── test_outputs.py     # pytest assertions

5. ⚙️ Configuration

Variable	Default	Description
CRAWL_COUNT	5000	Questions to crawl per round
GEN_WORKERS	16	Parallel threads for generation
VAL_WORKERS	8	Parallel threads for validation
VAL_TIMEOUT	300	Docker timeout per step (seconds)
MODEL_NAME	claude-opus-4-6	LLM model name

6. 📋 Requirements

• Python 3.10+
• Docker (for validation step)
• An OpenAI-compatible API endpoint
• StackOverflow API key (optional, increases rate limits)

7. 📄 License

MIT