HuggingFace AI Coding Tools — Agent Traces

This dataset rehydrates the benchmark results from davidkling/hf-coding-tools-dashboard into the JSONL session format consumed by the Hugging Face Agent Trace Viewer.

What's inside

• 5 sessions, one per (tool, model, effort, thinking) configuration
• 581 query → response turns total (≈1,162 events)
• Tools covered: claude_code, codex, cursor
• Models: claude-opus-4-6, claude-sonnet-4-6, composer-2, gpt-4.1, gpt-4.1-mini

Each row in the original results table maps to two JSONL events in a session:

1. A user turn with the benchmark query_text
2. An assistant turn with the model's response, plus a benchmark_metadata block carrying cost, latency, token counts, detected HF products, mentioned competitors, and the query's level/category.

Format

Sessions use the Claude Code JSONL schema, which the trace viewer auto-detects:

{"type":"user","message":{"role":"user","content":"..."},"uuid":"...","parentUuid":null,"sessionId":"...","timestamp":"..."}
{"type":"assistant","message":{"role":"assistant","model":"...","content":[{"type":"text","text":"..."}],"usage":{...}},"uuid":"...","parentUuid":"<user uuid>","sessionId":"...","timestamp":"...","benchmark_metadata":{...}}

UUIDs are deterministic (SHA-1 of session::role::row_id) so the parent/child chain is stable across regenerations.

File naming

{tool}__{model}__{effort}__{thinking}.jsonl

See manifest.json for the full session index.

Source

Generated from the results in davidkling/hf-coding-tools-dashboard.

Credits

Format adaptation originally pioneered by Clem Delangue in clem/hf-coding-tools-traces.