cloudsurf-software/CloudSurf-4B-FC-bfcl-results

CloudSurf-4B-FC — raw BFCL V4 result files

Raw, unmodified BFCL V4 evaluation outputs backing the leaderboard submission PR ShishirPatil/gorilla#1357 for CloudSurf-4B-FC (a google/gemma-4-E4B-it fine-tune, Apache-2.0).

Both sides are included: our tuned runs and the stock gemma-4-E4B-it baselines re-measured on the identical rig, so every number in the PR can be recomputed from primary files.

Whiskers are the min–max across the three runs on each side. Stock wins Irrelevance (84.4 vs 80.6) — the tune trades a little abstention for large gains on Multi-Turn, Web Search, and Memory.

Browsable scores

bfcl_v4_scores.csv (what the dataset viewer above shows) is a tidy long-format extract of every run's score/data_overall.csv: one row per (run, category), 25 categories per run, all 11 runs — the 3 tuned seeds, the 5 stock runs (3 thinking + 2 no-think), and the 3 w0-verify fairness arms (champion re-eval, champion + strip, stock + strip). Columns: run, model, variant, seed, category, accuracy_pct.

Contents

Everything is in gorilla-pr-results-cloudsurf4bfc.tgz (193 MB, 2,310 files):

Directory	What it is
e4b-rev1/	CloudSurf-4B-FC seed 42 — full 22-category run (result/ + score/)
e4b-rev2/	seed 7 — full run
e4b-rev3/	seed 11 — full run (released champion checkpoint)
e4b-stocknoise/	stock gemma-4-E4B-it, 3 independent runs (the baseline band)
w0-verify/	same-box fairness pair: champion + stock re-run with and without the stray-closer strip, plus a best-checkpoint re-eval

Each run keeps BFCL's native layout: result/<model>/<category>/*_result.json (raw generations) and score/ (scored CSV/JSON), plus the repro-kit file recording harness commit, backend, and flags.

tar -xzf gorilla-pr-results-cloudsurf4bfc.tgz
sha256: 0e5ebff3e8ad57e1e57676b29211941690cf4f2e3ae622f36b39b198e01ff8bc

Settings (identical for tuned and stock)

• BFCL V4 harness pinned at f7cf735 (the commit the public board cites)
• Prompt mode, thinking ON, sglang backend, bf16, self-hosted 1×H100
• Full 22-category runs; Overall = unweighted mean over all subcategories

Headline numbers (recomputable from these files)

• As-registered: tuned 3-seed mean 55.73 (55.53 / 54.86 / 56.81) vs stock mean 34.81 ± 1.68
• Matched-variant frame (both sides with the stray-closer strip, same box): champion 54.91 vs stock 39.40
• The stray <tool_call|> closer strip in the PR raises stock E4B by +4.6 Overall (+13.7 multi-turn) — measured both sides; see w0-verify/

Known caveats (poison counts per category, eval-band spread, web-column error rates) are documented in the PR body and reproducible from the *_result.json files.

Contact

brady@cloudsurf.com (CloudSurf Software)