docs/v2-data-card.md

---
language:
- en
license: other
license_name: mixed-upstream-licenses
license_link: https://huggingface.co/datasets/TypeScript/FastBrain-Instruct/blob/main/LICENSE.md
task_categories:
- text-classification
- zero-shot-classification
task_ids:
- intent-classification
- multi-class-classification
tags:
- fastbrain
- tool-use
- function-calling
- intent-classification
- multi-label
- synthetic
- aggregated
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
- WillHeld/mtop
- WillHeld/top_v2
- prem-research/Funcdex-MT-Function-Calling
size_categories:
- 100K<n<1M
configs:
- config_name: default
  default: true
  data_files:
  - split: train
    path: train.parquet
  - split: test
    path: test.parquet
- config_name: choice_v2
  data_files:
  - split: train
    path: data/choice_v2/train.parquet
  - split: validation
    path: data/choice_v2/validation.parquet
  - split: test
    path: data/choice_v2/test.parquet
- config_name: semantic_v2
  data_files:
  - split: train
    path: data/semantic_v2/train.parquet
  - split: validation
    path: data/semantic_v2/validation.parquet
  - split: test
    path: data/semantic_v2/test.parquet
- config_name: tool_context_v2
  data_files:
  - split: train
    path: data/tool_context_v2/train.parquet
  - split: validation
    path: data/tool_context_v2/validation.parquet
  - split: test
    path: data/tool_context_v2/test.parquet
- config_name: nlupp_hotels_v2
  data_files:
  - split: train
    path: data/nlupp_hotels_v2/train.parquet
  - split: validation
    path: data/nlupp_hotels_v2/validation.parquet
  - split: test
    path: data/nlupp_hotels_v2/test.parquet
- config_name: semantic_original_labels
  data_files:
  - split: train
    path: data/semantic_original_labels/train.parquet
  - split: validation
    path: data/semantic_original_labels/validation.parquet
  - split: test
    path: data/semantic_original_labels/test.parquet
---

# FastBrain-Instruct

FastBrain-Instruct aggregates public intent and tool-selection data for decision heads. The **2026-10-14 v2** release adds semantic boundary pairs, richer tool context and a separate multi-label task while preserving the original files.

Load the integrated release explicitly:

```python
from datasets import load_dataset

ds = load_dataset("TypeScript/FastBrain-Instruct", "choice_v2")
print(ds["train"][0])
```

## Configurations

| Config | Train | Validation | Test | Use |
|---|---:|---:|---:|---|
| `default` | 97,450 | — | 10,828 | Frozen original release |
| `choice_v2` | 116,440 | 8,206 | 24,639 | Integrated choice corpus: legacy replay + semantic pairs + tool context |
| `semantic_v2` | 26,976 | 7,230 | 12,228 | MTOP/TOPv2 intent boundaries and paired candidate-relative no-match |
| `tool_context_v2` | 26,241 | 976 | 1,583 | Funcdex and repaired Nemotron observed next-tool steps |
| `nlupp_hotels_v2` | 765 | 97 | 98 | Original multi-label intent modules and slots; separate task format |
| `semantic_original_labels` | 21,966 | 3,151 | 9,131 | Original labels for complex/ambiguous/unsupported/low-support cases; auxiliary only |

`semantic_v2` and `tool_context_v2` are **overlapping views of `choice_v2`**, not additional examples to concatenate with it. `nlupp_hotels_v2` and `semantic_original_labels` use their own schemas. Calling `load_dataset("TypeScript/FastBrain-Instruct")` still loads `default`.

The original `train.parquet`, `test.parquet`, `train.jsonl` and `test.jsonl` remain unchanged from [the frozen baseline](https://huggingface.co/datasets/TypeScript/FastBrain-Instruct/tree/a84ea0337a424e287dc6f36f6b9fa320fb920fb0). All 10,828 original test rows also appear in `choice_v2/test` with their original seven core fields intact. The original data card is [archived here](docs/legacy-data-card.md); its historical claims should be read together with the new audit below.

## What changed

| Main training component | Rows |
|---|---:|
| Retained legacy replay | 63,223 |
| MTOP/TOPv2 semantic pairs | 26,976 |
| Funcdex | 4,010 |
| Repaired Nemotron | 22,231 |

- **Semantic coverage:** 68 retained intent classes after alias consolidation, split/template isolation, label screening and a minimum of 20 independent training requests per class. Each selected request has one positive and one answer-removed no-match example. Candidate budgets target 16/32/64 ordinary operations and shrink when conflict exclusions require it; `__no_match__` is counted separately. This release does not provide a 151- or 256-intent natural-language catalog.
- **Tool context:** preserve the full available pre-answer prefix for Funcdex; restore Nemotron system policy and exclude steps with known missing earlier history or non-system truncation. Tools remain **observed expert-action labels**; schema validity and context recovery do not establish that every other tool is invalid. No text-only/no-call record is relabeled as no-match.
- **Holdout isolation:** the audit identified 2,687 raw Nemotron conversations shared by the original train and test sets. The v2 rebuild blocks 3,286 historical held-out parents before extracting training steps. The frozen `default` retains its historical split limitations.
- **Multi-label preservation:** NLU++ hotels keeps all annotated intent modules and slots. Only 576 of its 765 train rows are eligible for intent training; empty-intent context fragments have `use_for_intent_training=false` and no fabricated negative labels.

## Record format

Choice configs retain `id`, `domain`, `category`, `query`, `target`, `choices` (a list of name/description objects), and `source`. Added fields record `parent_id`, `source_family`, `source_repo`, `source_revision`, `source_split`, `source_row_id`, `task_type`, `label_status`, `is_rejection`, `candidate_count`, `catalog_id`, `provenance_json`, `context_json`, and `tools_json`.

`provenance_json` contains the source file and row, construction method, group, and applicable license for new rows. For tool records, `query` renders the structured `context_json`. Funcdex retains full original schemas in `tools_json`; Nemotron retains only names, descriptions and required-parameter names because the intermediate source discarded the rest.

NLU++ uses `text`, `intents`, `slots_json`, source/fold identifiers, `use_for_intent_training`, `head_questions_json` and `head_expected_json`. The optional head fields contain 40 independent boolean module questions. Filter on `use_for_intent_training` before intent training. Auxiliary semantic rows retain source labels and parse trees and explicitly set `use_for_single_choice=false`.

## Curation and use

[Curation recipe and split policy](docs/curation-v2.md) · [Source revisions and file hashes](audit/source_manifest.json) · [Dataset counts and hashes](audit/release_summary.json) · [Validation report](audit/validation.json) · [Semantic catalog](catalogs/semantic_v2.json) · [Source exclusions](curation/excluded_source_families.json)

Use `source`, `source_family`, `task_type` and `provenance_json.group` to make explicit ablations. The paired semantic data has an artificial 1:1 present/absent ratio within that view; this is not a deployment class prior. The release does not establish model improvement or preservation of an existing head's capabilities; those require training and evaluation.

The source-family exclusions from the original corpus are retained. Additional lexical checks use the exact cached files recorded in the exclusion manifest. They do not certify absence of paraphrases or unobserved source ancestry. Curation uses assistant ontology review, deterministic rules and targeted training-example inspection, **not exhaustive human annotation**. Ambiguous single-choice cases are retained only in the explicitly auxiliary view when possible.

## Attribution and licenses

This is an aggregation and adaptation of upstream data. MTOP and TOPv2 are **CC BY-SA 4.0**; NLU++ and the attributed Nemotron component are **CC BY 4.0**; Funcdex declares **MIT**. Legacy records retain their original source terms, including SGD's CC BY-SA 4.0. The collection is not licensed uniformly as MIT. See [license and attribution notices](LICENSE.md), the pinned source manifest and the archived legacy card.