RANK.AI / PUBLISHED METHOD

BFCL v4 tool-use configuration ranking reviewed 2026-04-13

Ranks exact model-plus-inference configurations by official UC Berkeley BFCL overall accuracy.

Version: bfcl-v4-2026-04-13-official-v1-bfcl-tool-use-configurations-v1
Published: Jul 27, 2026
Live boards: 1
Method ID: bfcl-tool-use-configurations

SOURCE CONTRACT

Where the facts come from

Source: bfcl-official
License Url: https://github.com/ShishirPatil/gorilla/blob/main/LICENSE ↗

RANKING RULE

How the order is produced

official BFCL Overall Acc descending

REFRESH POLICY

When the method reruns

daily

LIVE BOARD COVERAGE

Boards using this version

Only live, quality-gated boards appear. Each board opens on the exact published table and evidence path.

Board	Primary measure	Rows	Confidence	Freshness	Board version	Cadence	Evidence
BFCL Tool-Use Configurationsbfcl-tool-use-configurations	bfcl-overall-accuracy	109	100%	current	bfcl-v4-2026-04-13-official-v1-bfcl-tool-use-configurations-v1	daily	Open board →

VERSIONED CONTRACT

Complete measurement ledger

These fields are stored with the methodology version. A changed contract must publish a new version before comparable movement can resume.

Source

bfcl-official

License

Apache-2.0

Evidence

Every ranking component links to the exact normalized observation and the immutable source-artifact record.

Identity

Rows retain BFCL's exact configuration label. Function-calling, prompt, thinking, and unspecified modes remain distinct systems.

Universe

All 109 exact model-plus-inference configurations in the SHA-pinned official BFCL leaderboard capture.

Benchmark

Berkeley Function Calling Leaderboard

License Url

https://github.com/ShishirPatil/gorilla/blob/main/LICENSE ↗

Repository

https://github.com/ShishirPatil/gorilla/tree/main/berkeley-function-call-leaderboard ↗

Versioning

The official CSV URL, SHA-256 digest, exact row count, complete header, ordering, and reviewed date are code-pinned.

Limitations

This is an independent tool-use benchmark family, not an absolute model score.
Rows compare model-plus-inference configurations, not base weights alone.
Total cost and latency are BFCL workload context, not normalized market prices.
Scores may not generalize beyond BFCL V4 tool-use tasks.
A changed upstream CSV is suppressed until a new capture is reviewed and pinned.

Missing Data

Publication fails unless every configuration has every required metric. Optional format-sensitivity values are omitted when BFCL reports N/A; no value is imputed.

Rank Formula

official BFCL Overall Acc descending

Tie Breakers

official BFCL upstream rank ascending
canonical entity slug ascending

Reviewed As Of

2026-04-13

Primary Metric

bfcl-overall-accuracy

Optional Metrics

bfcl-format-sensitivity-max-delta
bfcl-format-sensitivity-standard-deviation

Benchmark Version

Ranking Direction

descending

Required Metrics Per Entity

bfcl-overall-accuracy
bfcl-non-live-ast-accuracy
bfcl-live-accuracy
bfcl-multi-turn-accuracy
bfcl-web-search-accuracy
bfcl-memory-accuracy
bfcl-relevance-detection-accuracy
bfcl-irrelevance-detection-accuracy
bfcl-evaluation-total-cost-usd
bfcl-latency-mean-seconds
bfcl-latency-p95-seconds