Flownium/zeus-dataset

Zeus Training Mixture

A source-attributed collection of structured decision examples and execution-trace records. It combines the corrected 30,000-trace Zeus release with Typed Decisions and a curated, consistently parsed subset of ToolACE.

The default decisions configuration contains 81,592 examples covering answer selection, yes/no judgments, ordered scores, evidence use, tool selection, and next-step decisions. Every one of the 30,000 Zeus trace IDs is represented. The companion trace_records configuration contains 30,000 records with conversation context, tools, final responses when available, and recorded reasoning where permitted by the source eligibility flags.

Sources

Source	Included decision examples, all splits	Original dataset
Zeus traces	60,557	Flownium/zeus-30k-trace-dataset
Typed Decisions	8,000	LocalLLaMA/typed-decisions
ToolACE	13,035	Team-ACE/ToolACE

Zeus contributes coding, mathematics, writing, document processing, tool use, clarification, and failure-recovery tasks. Its traces contain 140 task families; parameter variations are distinct inputs within those families.

Typed Decisions contributes 1,600 cases across agent-trace observability, customer service, invoice processing, and security incidents. Each case supplies five questions and teacher-derived probability targets. Only its all configuration is used, avoiding duplication of the separately packaged workflow subsets.

ToolACE contributes next-step function-selection examples from its published synthetic conversations. Targets identify the functions used by the next recorded assistant turn, or a no-function action. Multiple functions are represented as a set; their argument values and call order are not decision targets. The build supports the source's declared JSON function lists and bracketed call notation. Other tool-description formats, empty catalogs, empty assistant turns, and ambiguous call-like prose are excluded and logged.

Configurations and splits

Decision split	Zeus traces	Typed Decisions	ToolACE	Total
train	48,816	6,000	11,676	66,492
validation	2,908	0	657	3,565
gate_fit	2,949	0	0	2,949
calibration	2,944	0	0	2,944
test	2,940	2,000	702	5,642

Zeus keeps its original family-separated train, validation, gate_fit, calibration, and test assignments. Typed Decisions keeps its original train/test assignments. ToolACE has no upstream held-out split; this release assigns approximately 90%/5%/5% to train/validation/test using a deterministic hash of each declared tool catalog. Conversations with the same catalog stay together.

trace_records retains the original Zeus splits. It includes 26,293 source-eligible positive response records. One trace has no final response and is retained as an explicitly ineligible unfinished record. Record counts across configurations overlap and should not be added to count independent tasks.

Record format

Both configurations use the same flat Parquet columns. Variable-shaped content is stored as JSON strings for stable loading.

Field	Meaning
example_id	Unique, source-prefixed example identifier.
source, source_repo, source_revision, source_id	Source identity and pinned revision.
source_group_id, source_split, split	Grouping and original/current split assignments.
category, task_type, label_status	Task category, output type, and label provenance.
input_json	Conversation prefix, tools, question/options for decisions, and previously emitted reasoning when available.
target_json	Correct option IDs, soft probabilities where supplied, or a response target. Null for the unfinished trace.
provenance_json	Source references, qualification details, and transformation metadata; excluded from model input.
eligible_positive_example, exclusion_reason	Whether the row is eligible as a positive example under its recorded source checks.
input_sha256	Fingerprint of normalized input, ignoring option order and transport-only call IDs.
sampling_weight	Optional source-balanced training weight; zero for held-out or ineligible records.

Decision targets use correct_option_ids. Typed Decisions also supplies a probabilities map and, where available, expected_score or probability_true. Probabilities are renormalized only to remove small rounding drift; original targets remain in provenance. Hard labels are not assigned invented confidence values.

Reasoning is copied from recorded Zeus API emissions. Decision inputs contain only reasoning emitted before the decision prefix boundary. Current-turn response reasoning is a target, not an input. Reasoning marked partly ineligible by the source is omitted. External examples have no fabricated reasoning. Source qualification and response-eligibility flags are preserved.

Training-mixture weights

The files store unique retained examples without physically repeating rows. An optional initial sampler assigns 80% Zeus / 10% Typed Decisions / 10% ToolACE probability mass within the eligible decision training split. Each row's sampling_weight distributes its source's mass uniformly. Loading the data normally does not apply these weights automatically; the consuming sampler must use them.

This ratio is a starting configuration, not a measured optimum. trace_records has its own uniform weights across eligible training responses and is a separate view, not part of the 80/10/10 decision ratio.

Loading

import json
from datasets import load_dataset

train = load_dataset("Flownium/zeus-dataset", "decisions", split="train")
train = train.filter(lambda row: row["eligible_positive_example"])
example = train[0]
inputs = json.loads(example["input_json"])
target = json.loads(example["target_json"])

traces = load_dataset("Flownium/zeus-dataset", "trace_records", split="train")

Use input_json for inputs and target_json for supervision. Source IDs, verification records, latent factors, label evidence, and later conversation turns are not input features. Held-out splits remain reserved for their stated purpose.

Preparation and quality checks

Source files are pinned and verified against repository content hashes. Preparation checks exact trace coverage, unique IDs, target membership, probability totals, causal prefix boundaries, split-group isolation, and exact normalized-input duplicates. Duplicate decisions retain the highest-priority held-out occurrence; conflicting targets for identical normalized inputs are excluded. This is exact duplicate filtering, not semantic near-duplicate detection.

Label quality retains each source's scope. Zeus uses its published reference checks and review qualifications. Typed Decisions labels measure agreement with an upstream teacher, rather than independently established truth. ToolACE targets follow source demonstrations; this preparation does not execute the tools or independently certify every action. Recorded reasoning may contain source errors even when a final answer is correct. These are dataset qualifications, not claims of perfect semantic accuracy.

Exclusion log summary (entries can identify a whole conversation or one decision):

• ambiguous_call_in_prose: 10 exclusions.
• duplicate_normalized_input: 3 exclusions.
• empty_assistant_turn: 5 exclusions.
• empty_function_list: 248 exclusions.
• missing_function_list_marker: 518 exclusions.

Provenance and reproducibility

Release date: 2026-10-07.

Pinned source revisions:

• Flownium/zeus-30k-trace-dataset at 68de5f7a9c24

• LocalLLaMA/typed-decisions at d0e2f0c42fef

• Team-ACE/ToolACE at 6bda777c88d2

• mixture_manifest.json: exact counts, source weights, transformations, and coverage.

• validation_report.json: validation results for the exported Parquet files.

• source_manifest.json: source file sizes and hashes.

• source_snapshots.json: pinned repository metadata.

• excluded_examples.jsonl: excluded source identifiers and reasons.

• file_manifest.json: sizes and hashes of the published release files.

• preparation/: conversion and release-preparation code with dependencies.

Attribution and licensing

The Zeus compilation and Flownium contributions follow this repository's CC BY 4.0 designation. Typed Decisions and ToolACE material retains its upstream Apache 2.0 terms. See LICENSE.md, NOTICE, and the included Apache 2.0 license.

ToolACE citation: Weiwen Liu et al., ToolACE: Winning the Points of LLM Function Calling. Source dataset links and pinned revisions above provide attribution for every component.