pyuvraj/bfcl-v1-jev-routing

BFCL V1 Jev Routing Pilot

250 frozen evaluation cases for selecting an assistant's next tool or choosing no function. This is a small, BFCL-derived routing pilot. It is not an official BFCL leaderboard benchmark, a complete function-calling evaluation, or a new Jev model.

The motivation is practical: can a small local decision model select the next capability for a personal assistant, leaving a chat model to fill arguments and explain results? This release makes the initial routing test inspectable and reproducible. It is a first 250-case pilot for that project, not a claim to have invented tool routing or Jev evaluation.

• Public article and interactive results
• Packaged conversion and selection tools
• Sanitized results and all recorded predictions
• Every original frozen case
• Selection manifest · selection review
• Release checksums

Results and decision latency

Seven existing decision systems were evaluated on the same 250 frozen cases. The figures below use the recorded predictions in artifacts/report.json. The long-term objective is efficient, long-context tool selection for on-device assistants; this experiment measures first-step selection on short inputs.

Overall results

Figure 1. Overall selection accuracy and median observed decision wall latency. Accuracy requires the reference function on 200 positive cases and accepts any non-tool action on 50 negative cases. Timings include the first recorded calls and reflect different precisions, native runtimes, and request boundaries. These measurements are exploratory; the hardware model is not disclosed. Download PNG · Download SVG

Selection accuracy by evaluation subset

Figure 2. Selection accuracy across 150 multiple-tool, 50 one-tool positive, and 50 no-function cases. Positive cases require the exact function; no-function scoring accepts no_tool, clarify, or cannot_answer. Correct arguments and tool execution are not scored. Download PNG · Download SVG

Accuracy versus decision latency

Figure 3. Overall selection accuracy versus median observed decision wall latency on the same 250 cases. One decision was recorded per case without a uniform warm-up or repeated-trial protocol. Backend, precision, kernels, and timing boundaries differ. These observations do not establish Mac performance, an architecture ranking, or an end-to-end assistant speedup. Latency is measured in milliseconds; lower is faster. Download PNG · Download SVG

The interactive report provides subset filters, mean/median timing views, and per-case predictions. Its references credit all evaluated checkpoints, scoring methods, and source datasets. Exact model revisions and runtime settings are also preserved in the results artifact.

Composition

There is one test split and no training or validation split.

Stratum	Cases	Offered functions	Expected outcome
multi_tool_call	150	2–4	One specific function
one_tool_call	50	1	The offered function
one_tool_no_call	50	1	No function
Total	250		200 positive / 50 negative

The source-category quotas are multiple_function 120, executable_multiple_function 30, simple 35, executable_simple 15, and relevance 50. The name relevance is from the pinned V1 source: it means that none of the offered functions should be called. Do not confuse it with similarly named categories in later BFCL versions.

All cases retain the three assistant disposition options no_tool, clarify, and cannot_answer, followed by the example's original function candidates. Their historical wording, order, and shortened descriptions are preserved exactly. Some wording was inherited from the personal-assistant harness; it is not an independently validated general policy.

What the labels mean

Primary tool-selection metric: on the 200 positive cases, the predicted action must be the exact gold_action_id function. On the 50 negative cases, any of no_tool, clarify, or cannot_answer counts as selecting no function. Invalid or failed outputs count as wrong. These three actions are collapsed only for this no-function outcome metric.

Exact-action diagnostic: compare the selected ID directly with gold_action_id. The frozen negative label is no_tool, but BFCL supplies no evidence that this is better than clarify or cannot_answer. Thus a lower exact-action score need not imply worse assistant behavior. This dataset cannot validate the choice among those three dispositions.

Report all three strata and the 200-positive/50-negative denominators separately. The 50 one-tool positive cases are easy controls; always choosing the offered function gets all 50 correct. Across the full split, choosing the first offered function gets 106/250; always choosing no function gets 50/250. These baselines are recorded in the selection manifest.

Load the viewer-friendly split

from datasets import load_dataset

ds = load_dataset("pyuvraj/bfcl-v1-jev-routing", split="test")
row = ds[0]
model_input = {
    "state": row["state"],
    "questions": {
        "next": {
            "type": "choice",
            "instructions": row["instruction"],
            "criteria": {o["id"]: o["description"] for o in row["options"]},
        }
    },
}

Only state, instruction, and options define the canonical decision input. Keep gold labels and original references out of the model prompt. Full schemas are archival information; the pilot routed on concise description cards, without parameter schemas. Adding full schemas creates a different input condition and should be reported separately.

Field	Meaning
id, source_id	Stable converted case ID and upstream source ID
stratum	One of the three groups above
user_request	Original BFCL request
state	Exact frozen Laya text state, including first-step context
instruction	Exact frozen choice question
options	Ordered list of id, label, description, kind; kind is tool or meta
gold_action_id	One derived exact-action target
gold_tool_names	Singleton function list for positives; empty for negatives
expected_route_kind	tool_call or no_function
full_schemas_json	JSON string preserving original heterogeneous function schemas
original_reference_json	JSON string preserving BFCL reference; "null" for V1 no-call category
source_category, source_row_number	Upstream category and one-based row
source_repository, source_revision	Pinned BFCL repository and commit
source_question_file, source_answer_file	Source-relative files; empty answer filename when absent
label_source	Reference-answer extraction or category-derived no-call label

json.loads recovers the two archival JSON-string columns. They remain strings so the Dataset Viewer does not infer incompatible nested types across different API schemas. The complete original case structure, including both Laya and MacJev payloads, is available unchanged under provenance/frozen/; it is deliberately excluded from automatic split loading.

Provenance and reproducibility

Source: Gorilla / BFCL V1 tag v1.0, commit 9df5c346ee0556c8a7cb09fd7206a39aadd904c2. The original data card dates the dataset February 26, 2024. The 250 selected cases originate from a 2,000-case conversion, not a newly collected personal-assistant dataset. No private user traces or screenshots were added.

The converter extracts reference function names; it parses executable reference strings as syntax and never executes them. Selection filters non-unique first actions, unsupported categories, conversion warnings and an explicit wrong-capability exclusion. An explicit 55-ID allowlist limits the V1 no-call pool. Within each category, SHA-256 ranking with seed bfcl-v1-jev-routing-250-v1 selects the quota while removing repeated normalized requests. The original selection review and complete rules are included. This is a selected pilot, not a representative random sample of all BFCL or assistant requests.

Frozen case SHA-256: 68868277c9f10c56706a5d8a1a79a78720582dd731fef414a4b7c4ae366cc602

Viewer split SHA-256: 052e289b4e8dd09b292aa8e812bb2947fdb85137d5738426147a04581c37e135

The parent conversion manifest records hashes of all upstream input files. release_manifest.json records hashes of the packaged files. The original cases, selection manifest, index and review are copied byte for byte.

After cloning this dataset repository, verify/rebuild its projection offline using only Python's standard library:

python3 tools/build_hf_release.py --input-dir provenance/frozen --output-dir . --check
python3 tools/build_hf_release.py --input-dir provenance/frozen --output-dir .cache/rebuilt

To reproduce the original conversion and selection from pinned BFCL source:

git clone --no-checkout https://github.com/ShishirPatil/gorilla.git .cache/bfcl-v1
git -C .cache/bfcl-v1 checkout 9df5c346ee0556c8a7cb09fd7206a39aadd904c2
python3 tools/convert_bfcl_v1.py --source-dir .cache/bfcl-v1/berkeley-function-call-leaderboard/data --output-dir .cache/bfcl-v1-converted
python3 tools/select_eval_subset.py --input .cache/bfcl-v1-converted/cases.jsonl --output .cache/reselected --review-file tools/bfcl_v1_selection_review.json
python3 -c 'import hashlib,pathlib; p=pathlib.Path(".cache/reselected/cases.jsonl"); assert hashlib.sha256(p.read_bytes()).hexdigest()=="68868277c9f10c56706a5d8a1a79a78720582dd731fef414a4b7c4ae366cc602"'

Intended use and limits

• Compare exact tool choice and coarse no-function decisions on identical frozen inputs.
• Inspect failures and test candidate-order or prompt changes as separately named experiments.
• Keep these public test cases out of training and threshold tuning when claiming held-out evaluation. BFCL may already occur in model pretraining; contamination status is unknown.
• This set does not score arguments, execution, permissions, browser control, memory, final answers, recovery after tool errors, or complete multi-step assistant tasks.
• Some positive requests could be answered without tools. Labels encode BFCL-conditioned routing rather than the uniquely best autonomous action.
• A 250-case pilot cannot establish general superiority or reliable personal-assistant behavior. Near-ceiling differences of a few examples should not be treated as decisive rankings.

The associated seven-model runs in the public article are exploratory latency measurements; the hardware model is not disclosed. Local-assistant motivation does not turn those results into Apple Silicon measurements. This dataset release makes no speed or end-to-end task-success claim. The separate artifacts/report.json contains sanitized aggregate results and per-case predictions; it is excluded from the test split. It records runtime limitations alongside the observed timings. Packaged code reproduces the data conversion and selection; model inference adapters are not included in this dataset release.

Related work

This release is an independently converted, frozen pilot and does not claim to be the first:

• BFCL / Gorilla supplies the underlying requests, function documentation and reference labels.
• Blobfish Typed Decision Bench: BFCL tool dispatch already evaluates typed choices on a 200-item BFCL V1-derived tool-dispatch set.
• OD1 reports BFCL-derived tool-selection comparisons with native and enlarged candidate catalogs, including Laya variants.
• LangWatch's Jev comparison includes tool routing among comparisons of Jev and small open decision models.
• JevBench studies broader typed decisions; SemIf provides direct-logit decision readouts.

Their inputs, model variants and scoring protocols differ. Their published scores are not interchangeable with results on this split. Jev and TypeSafe names identify related work; this project is independent and is not endorsed by their owners or the BFCL authors.

License and attribution

The BFCL-derived data and retained BFCL material use Apache-2.0; the complete upstream license is included as LICENSE and its data card as provenance/BFCL_DATA_CARD.md. See NOTICE for source attribution and the modifications made by this project. The original converter, selector, packaging code and original documentation use MIT, reproduced in licenses/CONVERTER_LICENSE.txt. Model weights and their licenses are separate and are not included here.

Maintainer: Pritish Yuvraj. Cite this dataset's pinned Hub revision and the upstream BFCL commit when reporting results, so subsequent revisions cannot silently change the test set.