pyuvraj/bfcl-v4-jev-routing

BFCL V4 Live Jev routing pilot

250 frozen prompts for capability selection, with saved decisions and probabilities from seven systems. Jev-style decision models choose the tool; a separate chat model supplies its arguments and asks for missing values before execution. This set scores one routing decision. It does not score function arguments, tool execution, final answers, or complete tasks, and is not an official BFCL V4 function-calling or agentic score.

Adaptation and evaluation: Pritish Yuvraj. Source data: Shishir Patil and the Gorilla / Berkeley Function Calling Leaderboard contributors. Source repository · BFCL leaderboard · GitHub release · Earlier V1 pilot.

Inspect or download

• Browse the 250-row test split.
• Download the offline chat-style viewer. Save it locally and open it in a browser; it embeds every prompt, all seven models' selections/probabilities, original schemas, and review material. The Hub file browser is not an HTML app host.
• Complete results report · Comparison CSV · All release files and hashes.
• Schema · Usage and exact rerun commands · Credits and checkpoint pins.

from datasets import load_dataset

ds = load_dataset("pyuvraj/bfcl-v4-jev-routing", "default", split="test")
assert len(ds) == 250

row = ds[0]
# Only these fields are inference inputs. Preserve option order and descriptions.
payload = {
    "state": row["state"],
    "questions": {"next": {
        "type": "choice",
        "instructions": row["instruction"],
        "criteria": {o["id"]: o["description"] for o in row["options"]},
    }},
}
# Send payload through the model's native adapter, then score its returned ID.

Prevent gold leakage: never serialize the whole dataset row to the model. gold_*, expected_route_kind, original_reference_json, source fields, review notes, model logs, and full_schemas_json are inspection/scoring material. The published runs sent only the frozen state, decision instruction, and offered option descriptions. The full schemas are retained for provenance; they were not added to those inference payloads.

The YAML config explicitly selects only test.jsonl. Frozen provenance and model-selection JSONL files are downloadable artifacts, not additional dataset splits. frozen_cases.jsonl is the byte-identical canonical evaluation file; test.jsonl is its flat Dataset Viewer projection.

Composition, source and selection

BFCL Live category	Stratum	Cases	Target
live_multiple	multi_tool_call	150	One relevant function among 2–4 offered tools
live_simple	one_tool_call	50	The one relevant offered function
live_irrelevance	one_tool_no_call	50	No applicable offered function

These three source categories contain 2,195 rows (1,053 / 258 / 884). They originated before V4 and remain in its corpus. This selection excludes parallel calls, multi-turn execution, web search, memory, and other V4 task families. It is a curated pilot, not a random representative sample of all BFCL V4.

Source commit: f7cf7359b7ac615a0b294831c5ba2bc95ee4a000. Frozen canonical case SHA-256: 2f7b2ad2d3f0d2610ce303fb3c10e7bdd00d03e4e2f9a7a826bf93293ebcb2e8. Seed: bfcl-v4-live-routing-250-20260929. Original source file hashes, row IDs, positive references, schemas, semantic-review reasons, all candidate dispositions, exclusions, and diversity checks are provided in the frozen_* files.

Accepted candidates were reviewed using the complete request, descriptions, schemas and positive reference function before evaluation. Missing argument values alone do not make a tool irrelevant: argument clarification is the chat model's responsibility. Negatives require a capability mismatch or a request needing none of the offered functions. Review is agent-assisted and published for inspection; residual label ambiguity is possible.

Within each stratum, selection ranks accepted rows by SHA256(seed + NUL + source_id) and caps normalized request, complete unordered tool catalog, and category + Live middle-ID family. The family cap is a conservative diversity heuristic, not a documented upstream universal family definition.

V1 request overlap is zero under both exact and normalized comparison with all 2,000 V1 source rows and the frozen V1 250. Normalization uses Unicode NFKC, case folding and Unicode word tokens joined with spaces. Four additional high-similarity candidates were explicitly excluded. This is not a guarantee against every semantic paraphrase or model training contamination.

V4 preserves full original tool descriptions with zero card overrides and zero automatic truncations. Frozen state text collapses request whitespace; user_request retains original source text. V4 uses versioned neutral meta cards (bfcl-v4-neutral-meta-cards/v1); V1 used compact cards and meta descriptions containing blanket capability limitations. Score differences between these two curated sets do not isolate a benchmark-version effect.

Requests include English, Chinese, Vietnamese and Polish; cards/instructions are predominantly English. For example, live_multiple_7-3-2 is Chinese, live_irrelevance_109-6-0 is Vietnamese and live_irrelevance_484-138-1 is Polish. This small mix does not support per-language rankings.

Metrics and recorded results

• Exact right tool: correct function ID on the 200 positive cases.
• Routing accuracy (primary): correct function on positives; any of no_tool, clarify, cannot_answer on 50 negatives.
• Exact action: exact function on positives and canonical no_tool on negatives. That negative meta target is a convention; BFCL irrelevance does not determine the best no-function disposition.
• Mean correct-route probability: gold function probability on positives and summed probability of the three meta actions on negatives, averaged over all 250 cases. This is option-conditional probability, not calibrated confidence or comparable raw logits.

All models completed 250/250, with zero request errors or retries. No model was tuned on this set. Checkpoint revisions and native scorer settings match the V1 runs; inputs use each model's own published prompt/readout adapter. All native input preflights passed without truncation.

Model	Approx. parameters	Exact right tool (200)	Routing (250)	Exact action (250)	Mean latency	Median	P95	Mean correct-route probability
Laya typed-decisions	~421M	186/200 (93.0%)	224/250 (89.6%)	192/250 (76.8%)	21.5 ms	17.1 ms	30.6 ms	61.1%
Kev-0.8B	~0.8B	180/200 (90.0%)	230/250 (92.0%)	199/250 (79.6%)	58.0 ms	57.3 ms	61.6 ms	59.2%
Bespoke Nimble 9B	~9B	200/200 (100.0%)	249/250 (99.6%)	222/250 (88.8%)	57.8 ms	47.6 ms	67.6 ms	98.2%
SemIf / frozen Qwen3.5-4B	~4B	196/200 (98.0%)	244/250 (97.6%)	229/250 (91.6%)	57.3 ms	50.3 ms	61.3 ms	91.9%
Rizzo Flow 4B Q8_0	~4B	195/200 (97.5%)	241/250 (96.4%)	232/250 (92.8%)	41.8 ms	40.4 ms	52.5 ms	92.6%
Von 1.3, chains off	~395M	53/200 (26.5%)	103/250 (41.2%)	56/250 (22.4%)	13.0 ms	12.9 ms	13.8 ms	38.8%
NanoJev unified-games-v1	~0.6B	7/200 (3.5%)	56/250 (22.4%)	35/250 (14.0%)	28.8 ms	25.9 ms	34.7 ms	30.3%

The low Von and NanoJev results are retained. Von selected no function on all 50 negatives but frequently rejected offered capabilities on positives. An independent NanoJev check found no adapter/option mapping defect in the checked paths. Its game-trained checkpoint is evaluated out of domain. SemIf is a scoring method over frozen Qwen weights, Rizzo uses Q8_0 GGUF weights, and Von chains are disabled. These results do not establish the cause of either poor score.

Charts

The probability chart averages the correct function probability on the 200 positives; the table's correct-route probability also includes the 50 negatives. SVG versions and exact metric definitions are in CHARTS.md and chart_metrics.json.

Timing and limitations

Runs used NVIDIA H100 80GB GPUs on vp-dgx-59 (dgxh100-059, driver 570.195.03), one model on each of GPUs 0–6. Models ran concurrently on separate GPUs; requests within each model were sequential. Parameter counts are rounded published backbone sizes; adapters/decision heads can add parameters.

All reported table latencies use saved wall_latency_ms over 250 responses, without outlier removal. P95 interpolates at 0.95 × (n − 1). HTTP and native direct scorers have different timing boundaries and runtimes/precisions. Warmup differs: Nimble/SemIf resumed scored two-case smoke runs in new processes; Rizzo/NanoJev resumed scored smoke on warm servers; Laya/Kev/Von used separate smoke requests before the scored run. Cold overhead present in saved rows is retained. Node concurrency can affect timings. These are descriptive H100 response times, not controlled throughput or laptop measurements.

The public prompts may have appeared in training. The set is short-context and single-step, contains curated rather than random cases, and measures only routing among the offered cards. Avoid treating 100% on this set as general agent reliability. Keep these prompts frozen for evaluation and use separate development cases for tuning.

Files, reproducibility and licensing

Seven original decision logs plus original .meta.json / .summary.json sidecars are retained byte for byte as results__bfcl_v4_*. Charts, metrics, the report, CSV, and the original run manifest accompany them. Complete runtime consoles, assets/library hashes, package snapshots, native preflights, and independent validation remain in GitHub runtime evidence; they are linked, not duplicated here.

See USAGE.md for loading a local JSONL, safe payload construction, hash checks, source reconstruction and exact per-model inference commands. Rebuilding the canonical evaluation requires the GitHub repository and pinned BFCL source checkouts; the flat HF clone is a release artifact, not a standalone inference checkout.

Adapted BFCL data retain Apache 2.0 licensing: LICENSE, upstream license copy, NOTICE. Original adaptation/evaluation/viewer code is MIT: CODE_LICENSE. Model weights and native implementations retain their own upstream licenses. CREDITS.md attributes every model, implementation and checkpoint revision.