willykeenan/agent-handoff-benchmark

Agent Handoff Benchmark

A text-free evaluation dataset for studying whether partial classifier output preserves a downstream decision. It packages the public Waggle/Kea BANKING77 reference artifacts as three directly loadable tables: intent predictions, 77-class probability vectors, and decision/certificate outcomes under 12 fixed synthetic cost policies.

William Keenan created the classifier experiment, derived predictions, decision-sufficiency experiment and this packaging. BANKING77 was created by PolyAI-LDN/Casanueva et al.; this is a derived evaluation dataset, not a new collection of banking queries. No source query text is included.

Companion classifier: willykeenan/banking-intent-classifier.

Load the dataset

Install the tested loader dependencies with Python 3.12:

python -m pip install -r requirements.txt

from datasets import load_dataset

repo = "willykeenan/agent-handoff-benchmark"
predictions = load_dataset(repo, "predictions", split="test")  # 3,050 rows
vectors = load_dataset(repo, "vectors", split="test")          # 3,050 rows
decisions = load_dataset(repo, "decisions", split="test")      # 36,600 rows

print(predictions[0]["trueIntent"], predictions[0]["predictedIntent"])
probabilities = [x / vectors[0]["probabilityScale"]
                 for x in vectors[0]["probabilities"]]

The decision table provides convenient scalar fields and retains the complete original record as a JSON string to avoid ambiguous, heterogeneous nested schemas across loading libraries:

import json
record = decisions[0]
original = json.loads(record["payload_json"])
certificate = original["adaptive"]["certificate"]
print(record["case_id"], record["policy_id"], certificate["disposition"])

Before publication, or from a downloaded snapshot, load a table directly:

local = load_dataset("parquet", data_files={"test": "data/vectors/test.parquet"}, split="test")

Configurations and schema

Configuration	Rows	One row represents	Main fields
predictions	3,050	One cleaned, train-disjoint BANKING77 test case	caseId, trueIntent, predictedIntent, wordOnlyIntent, confidencePpm, top3, sourceIndex, sourceSplit, typoEligible, textIncluded
vectors	3,050	The frozen 77-class prediction for that case	caseId, vectorId, modelId, labelIds, probabilities, probabilityScale, trueIntent, sourceIndex, sourceSplit, textIncluded, schemaVersion
decisions	36,600	One case under one of 12 synthetic policies	case_id, vector_id, policy_id, action/disposition columns, revealed_components, authority_granted, payload_json

All three configurations expose only a test split. They are related views, not independent sample pools: each case has one vector and 12 decision rows. Join predictions.caseId to vectors.caseId and decisions.case_id; join vectors.vectorId to decisions.vector_id. Every case-policy pair is unique. The 36,600 decisions are not 36,600 independent banking examples.

Vector probabilities are nonnegative integers in label order, with 77 values summing to probabilityScale = 1_000_000; divide by that scale before probability calculations. top3 is a list of intent/probability objects from the original prediction artifact. sourceIndex points to the pinned upstream test file; caseId is derived from normalized text. Text exclusion does not make records anonymous or prevent linkage to the public upstream source.

Decision scalar columns include adaptive_disposition, adaptive_action_id, qualification_disposition, restricted_disposition, restricted_action_id, full_vector_action_id, naive_top1_action_id, safe_k1_disposition, and safe_k3_disposition. Action IDs may be null when a control refuses to choose. revealed_components counts probabilities in the adaptive certificate. authority_granted is always false: a mathematical certificate grants no permission to execute an action. payload_json preserves every original field and value, including policies' expected costs, alternative controls, certificate bounds, qualification, byte counts and causal identifiers.

Provenance and construction

• Public source repository: willykeenan/waggle-kea, pinned to commit 54041045c82953cbbabac155b150a0f7b7ed5603.
• Upstream BANKING77: PolyAI-LDN/task-specific-datasets, pinned to commit 57ec275d8078af65b7731c2a98be812d844a6d6b; file digests appear in SOURCE.json.
• The existing protocol normalizes text to audit duplicates and ambiguity, retains 9,971 unique unambiguous training cases, and evaluates on 3,050 unique unambiguous test cases after excluding normalized train/test overlap.
• A fixed word-and-character TF-IDF / log-loss SGD classifier generates the predictions. The downstream study uses 12 deterministic, output-blind, four-action zero/one-cost policy matrices. These policies are synthetic and do not encode real bank procedures or customer-outcome labels.
• This package copies allowlisted public artifacts; it does not retrain models, generate cases, reassign labels, alter splits or recompute experimental results. Parquet preserves the prediction/vector records. The decisions table adds convenient projections while keeping the original row text in payload_json.

Original JSONL/gzip, policies, configuration, protocols, reference results and checksum manifests are retained under evidence/. SOURCE_FILES.json maps each copied file to its source path, SHA-256 and byte count. VALIDATION.json records row counts, schemas, exact round trips and relational checks. MANIFEST.sha256 pins this package's files. The build script verifies source bytes against the pinned Git commit and the original checksum manifests before packaging.

To reproduce the data tables from the public repository, install the supplied dependencies and run these commands from a downloaded copy of this package:

git clone https://github.com/willykeenan/waggle-kea.git source
git -C source checkout 54041045c82953cbbabac155b150a0f7b7ed5603
python scripts/build_dataset.py --source source --output reproduced-data

This rebuilds the tables and provenance checks, not the original experiments. The card and package documentation are maintained separately. The validator requires ordinary Python execution with assertions enabled (no -O).

Published results and limitations

The existing classifier comparison reports macro-F1 0.911858 for the word-and-character model versus 0.891545 for its matched word-only control. This comparison is exploratory and test-informed, not a confirmatory or state-of-the-art benchmark. See the retained classifier run.

The prospective secondary decision-sufficiency study qualified 35,013 of 36,600 decisions (95.66%), with zero qualified-decision mismatches against the full-vector reference. Its preregistered primary hypothesis was not supported: coverage improved by 6.65 percentage points over safe k=1, below the frozen 10-point threshold (H0_RETAINED). A simpler informed expected-cost summary was smaller than the certificate. These are retained results from the original evaluation and result report, not new measurements made for this package.

• This is test-only evaluation data. Do not train or tune on these cases and then report performance on the same cases as held-out evidence. Creating a different row split does not remove dependence among the 12 policies per case.
• It supports analysis of frozen predictions, uncertainty and decision preservation. It cannot train or evaluate a text encoder by itself because query text is absent.
• The guarantee is bounded to the frozen probability vectors, exact finite cost policies, specified verifier and omitted-mass model. It does not establish correct real-world customer outcomes or guarantee classifier correctness.
• No demographic fairness, privacy guarantee, natural-noise robustness, cross-domain generalization, production suitability, autonomous authority, semantic compression, universal agent communication, or overall token/cost/ energy advantage is claimed.
• No private institutional/customer dataset, model prompts, credentials, or personal local artifacts are included. Public-source linkability remains.

License and citation

Dataset packaging and derived evaluation contributions are provided under CC BY 4.0, with upstream BANKING77 attribution retained. Preserved source files keep their original notices; source code and the packaging script are MIT licensed. See LICENSE, source code license, and third-party notice. Please credit both the upstream dataset and the derived work, link the licenses, and identify any changes. Neither PolyAI nor the paper authors endorse this work.

Casanueva, I., Temcinas, T., Gerz, D., Henderson, M., and Vulic, I. (2020). Efficient Intent Detection with Dual Sentence Encoders. Proceedings of the 2nd Workshop on NLP for ConvAI.

For the derived benchmark, cite William Keenan's Waggle + Kea repository and the pinned commit above; CITATION.cff supplies metadata.