sup2ch/wex-agent-routing-v1

Wex Agent Routing v1

A small synthetic research dataset for selecting an agent's next repository inspection tool or abstaining, using only the visible task and retained observations. Tasks are primarily Russian; source identifiers and tool descriptions are often English.

Companion experimental model: https://huggingface.co/sup2ch/wex-v0.1.

Contents and splits

Split	Worlds / rows	Select	Abstain	Provisional HOLD
train	128	37	91	2
validation	64	23	41	1
test	64	20	44	0

There is one evaluated current-prefix cut per world. World IDs do not overlap between splits; mechanism families recur intentionally. This is not a mechanism-disjoint or real-user benchmark. The historical validation split participated in model comparison. The test split is the separate fresh64 check from 2026-10-07; it is now public and closed. Do not train, tune thresholds, select models or create derivative training examples from test labels and then claim independent evaluation on that same test.

Schema and use

Each row has id, case_id, world_id, mechanism_family, user_message, observation_count, decision, preferred_option_id, acceptable_option_ids, review_status, human_gold, request_json, and label_json.

request_json and label_json preserve the exact original objects as JSON strings. This avoids imposing a fixed Arrow schema on heterogeneous retained tool observations. The native request includes the task, offered options, applicability mask and retained observations. The label includes acceptable/useful choices, reasons, citations and argument witnesses. These witnesses explain labels; the model does not generate arguments.

import json
from datasets import load_dataset
ds = load_dataset("sup2ch/wex-agent-routing-v1")
request = json.loads(ds["train"][0]["request_json"])
label = json.loads(ds["train"][0]["label_json"])

Offered tools are list_directory, search_files, search_content, and read_file. abstain means no offered tool is judged a useful next step; it can include a state with enough information to answer. It does not necessarily mean failure or refusal. The original model accepts a dynamic option list, but this dataset covers this fixed four-tool catalog only. Generalization to arbitrary tool schemas is unproven.

Creation and label provenance

The tasks, synthetic repository worlds and histories were LLM-authored. Helpers mechanically constructed bounded reads/search matches and serialized JSON; examples are not harvested private LAH sessions. Expert labels were authored from visible current prefixes, without executing repository tools or reading future suffixes or predictions. All labels are machine expert proxies (human_gold: false), not human ground truth.

TRAIN and validation used separate model-blind expert contexts. Exact model identity is not embedded in the exported per-row labels. Their original review notes, including display limitations and retained reviewer context, are in provenance.json. The fresh test author is recorded as an existing gpt-6-astra author, and its reviewer as gpt-6-astra with medium reasoning in fresh context; the same model family therefore does not establish independence from shared teacher biases.

Two TRAIN and one validation HOLD labels remain in place. Published all-row metrics include them. No post-result relabeling or removal was done for this release.

Historical results and limitations

The companion semantic128 model scored 55/64 (85.94%) on validation and 54/64 (84.38%) on fresh test. On fresh test it abstained incorrectly in 5/20 (25%) select states, failing the original <=10% false-abstention criterion. A later candidate scored 44/64 (68.75%), with zero gains and ten losses against the reference, and was rejected. See evaluation.json for the original paired summaries and criteria.

Accuracy accepts any labeled acceptable tool; an abstain prediction is correct on an abstain label. These are cut-level proxy measurements, not whole-task completion, time saved, production reliability or calibrated confidence. The small, imbalanced, teacher-authored dataset cannot establish broad agent usefulness.

License

The exported synthetic data and accompanying files are released under MIT. No pretrained encoder weights are distributed in this dataset.

Project direction

Useful contributions include human review of ambiguous labels, genuinely new held-out worlds, whole-task agent comparisons and compact learned text representations. The public test must remain evaluation-only for comparable independent claims.

Русское описание: исследовательские синтетические примеры выбора следующего инструмента агента. Разметка машинная; практическая полезность модели ещё не доказана.