BFCL leaderboard traces

Every model response behind the Berkeley Function-Calling Leaderboard, with its grade — from the public BFCL-Result archive (one greedy rollout per question per model per release batch). Pick a batch and model, filter by question family / category / verdict, then click a row for the full trace: reasoning, tool calls, tool results, and environment state, in temporal order. Grading detail (error type, expected vs decoded call) comes from the score files; question + ground truth from the BFCL v4 benchmark data.

Custom SQL (advanced)

Prefilled with the SQL for the current selection (each Run refreshes it). The view bfcl is the whole index: one row per (batch, model, question) with columns batch, model, mode, family, category, id, passed, error_type, n_turns, n_steps, in_tok, out_tok, latency, has_reasoning, preview, rfile, roffset, sfile, soffset.

initializing…