docs/legacy-data-card.md

metadata

language:
  - en
license: other
license_name: mixed-upstream-licenses
task_categories:
  - text-classification
  - zero-shot-classification
task_ids:
  - intent-classification
  - multi-class-classification
tags:
  - fastbrain
  - system-1
  - non-autoregressive
  - function-calling
  - tool-use
  - routing
  - synthetic
  - composite
  - aggregation
  - decontaminated
source_datasets:
  - glaiveai/glaive-function-calling-v2
  - NousResearch/hermes-function-calling-v1
  - nvidia/Nemotron-SFT-Agentic-v2
  - MaziyarPanahi/AgentToolDecisions-180K
  - AnishJoshi/nl2bash-custom
  - burak29/git-natural-language-commands
  - AmazonScience/massive
  - google-research-datasets/schema_guided_dstc8
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.parquet
      - split: test
        path: test.parquet

⚡ FastBrain-Instruct: Aggregated Fine-Tuning Corpus for System 1 Decision Models

⚠️ Provenance: not an original dataset

FastBrain-Instruct is NOT written from scratch. It aggregates existing open datasets, reformats them into one choice/routing schema, removes finance content, deduplicates, and decontaminates against public decision benchmarks. All query text and labels come from the upstream sources below (except 25 maintainer-written DevOps templates).

Upstream source	Author	License	Rows (train / test)
Schema-Guided Dialogue (SGD)	Google Research (Rastogi et al., 2020)	CC BY-SA 4.0	34,455 / 3,889
Nemotron-SFT-Agentic-v2, via the tool_selection subset of AgentToolDecisions-180K	NVIDIA; transformed by Maziyar Panahi	CC BY 4.0	32,500 / 3,548
MASSIVE (en, via mteb/amazon_massive_intent)	Amazon Science (FitzGerald et al., 2022)	CC BY 4.0	14,264 / 1,531
Glaive Function Calling v2	Glaive AI	Apache 2.0	11,575 / 1,305
NL2Bash	Lin et al., 2018 (HF mirror by AnishJoshi)	see source	2,809 / 345
Hermes Function Calling v1	NousResearch	Apache 2.0	1,015 / 132
Git natural-language commands	burak29	see source	810 / 75
DevOps / terminal templates	maintainer	Apache 2.0	22 / 3

Glaive v2 and Nemotron-SFT-Agentic-v2 are themselves synthetically generated. Users must comply with each upstream license; SGD is CC BY-SA 4.0 (share-alike).

---

📌 What it is

Each row is a single decision: given a query (a user request, or for agent rows the recent conversation) and a list of candidate choices (name + description), pick the correct target. The format targets non-autoregressive "System 1" decision heads (e.g., the JointSchemaHead design used by Cloudflare Clef / Clef-Flash) that score candidates in one forward pass.

Two domains, balanced 50/50:

• code_mechanical — tool / function selection (Glaive, Hermes), multi-tool agent next-tool selection (Nemotron), terminal utility selection (NL2Bash), git subcommand selection, DevOps templates.
• daily_assistant — intent selection for voice-assistant and service dialogues (MASSIVE, SGD).

No model trained on this dataset is released, and no benchmark results are claimed.

🚫 Benchmark exclusion & decontamination

This dataset is intended for training models that will be compared on the benchmarks used in the Cloudflare Clef / Jev Decision Index table. Therefore:

1. No benchmark is used as a source. Excluded entirely: BFCL, ToolRet and all 35 of its source datasets (incl. APIGen/xLAM, ToolACE, ToolBench, Gorilla, MetaTool, API-Bank, …), API-Bank, When2Call (and derived data such as the When2Call part of AgentToolDecisions), BANKING77, CLINC150, BRIGHT, Amazon ESCI, PhishNChips.
2. Text-level decontamination. Every query (for agent rows: every user message) was checked against the texts of BFCL v3, ToolRet-Queries (35 subsets), When2Call (all files), API-Bank (all files), BANKING77 and CLINC150 (all configs). A row is removed on a normalized exact match or any shared 10-word n-gram. 2,674 rows were removed this way (e.g., Glaive: 899 ToolRet, 470 BFCL, 333 When2Call matches; SGD: 420 BFCL matches; MASSIVE: 212 CLINC150 matches).

Not checked at text level: Home appliances (no public copy found), BRIGHT, Amazon ESCI, PhishNChips (different domains from all sources used).

🔧 Processing

1. Extraction
   • Glaive v2: all 112,960 rows scanned; the user turn right before the first <functioncall> is the query; candidates are the tools listed in that sample's system prompt.
   • Hermes v1: first human turn; candidates are that sample's tools.
   • Nemotron agentic: conversation rendered as role: content lines (last 1,500 chars); candidates are all tools available to the agent (5–30).
   • NL2Bash: only single, non-compound commands (no pipes, ;, &&, substitutions, loops); target is the utility name; find capped at 1,500 rows.
   • SGD: user turns where a service's active intent changes; candidates are the intents of that service plus intents from other services, with SGD's own intent descriptions.
   • MASSIVE: all splits; description generated from the label name.
2. Candidates: at least 5 per row; when a sample has fewer own candidates, random distractors from the same source are added, excluding names that heavily overlap with the target (token Jaccard ≥ 0.5).
3. Finance removal: MASSIVE qa_currency, qa_stock; SGD Banks_* and Payment_* services.
4. Deduplication: normalized exact query duplicates removed; queries with conflicting labels dropped (1,410).
5. Balancing: the larger domain is downsampled (largest source first) to match the smaller one.
6. Split: random 90/10 over unique queries; no query appears in both splits.

📊 Statistics

	Train	Test	Total
Rows	97,450	10,828	108,278
code_mechanical	—	—	54,139 (50%)
daily_assistant	—	—	54,139 (50%)
Choices per row	5–30 (median 5)
Distinct targets			5,681

⚠️ Limitations

• Easy distractors in some sources: Glaive and NL2Bash rows usually have only one "own" candidate, so the other 4 are random and often unrelated.
• Near-duplicates remain: templated queries (Glaive, SGD) differing only in entities are kept; only exact duplicates were removed.
• Weak descriptions: MASSIVE descriptions are generated from label names (e.g., "User intention to play music"); Nemotron tool descriptions are kept as-is.
• NL2Bash skew: find dominates the terminal subset even after capping.
• Decontamination is lexical: paraphrased benchmark items would not be caught.

---

🛠️ Quickstart

from datasets import load_dataset

ds = load_dataset("TypeScript/FastBrain-Instruct")
print(ds["train"][0])

Field	Type	Description
id	string	Row id
domain	string	code_mechanical or daily_assistant
category	string	e.g. api_tool_routing, agent_tool_selection, terminal_command, service_dialogue
query	string	User request (or rendered recent conversation for agent rows)
target	string	Correct choice name
choices	list of {name, description}	Candidates, including target
source	string	Upstream source

⚖️ Upstream Attribution

• Google Research — Schema-Guided Dialogue (Rastogi et al., AAAI 2020), CC BY-SA 4.0.
• NVIDIA — Nemotron-SFT-Agentic-v2, CC BY 4.0; decision-format transformation by Maziyar Panahi (AgentToolDecisions-180K).
• Amazon Science — MASSIVE (FitzGerald et al., 2022), CC BY 4.0.
• Glaive AI — Glaive Function Calling v2.
• NousResearch — Hermes Function Calling v1.
• Lin et al. — NL2Bash (LREC 2018).
• burak29 — Git natural-language commands.

📜 Citation

@dataset{fastbrain_instruct_2026,
  author = {TypeScript},
  title = {FastBrain-Instruct: Aggregated Fine-Tuning Corpus for System 1 Decision Models},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {https://huggingface.co/datasets/TypeScript/FastBrain-Instruct}
}