Top-quality subset: 177 trajectories that are BOTH verified PASS AND contain full reasoning chains.

Key facts:

• 177/177 verified PASS (100% success rate)

• 177/177 with complete chain-of-thought reasoning

• Avg reasoning length: ~870 chars per trajectory

• Zero FAIL, zero NULL — only highest quality

• Ready for SFT of reasoning code agents

• Tools: run_python, read_file, write_file, list_files

• Format: JSONL, ATIF-v0.2

What makes this "premium":

Every single trajectory satisfies BOTH conditions:

1. The agent successfully completed the task (verified PASS by LLM judge)

2. The model explicitly reasoned before each tool call (full CoT)

This is the intersection of two normally-exclusive subsets, giving you the highest possible signal-to-noise ratio for training.

Perfect for:

- Fine-tuning code agents that reason before acting

- Research on agentic reasoning with guaranteed correctness

- Benchmarking reasoning quality on verified successful tasks

- SFT data where quality matters more than quantity