README.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
- dataset-repair
- fastbrain-v3
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: choice_v3
  data_files:
  - split: train
    path: data/choice_v3/train.parquet
  - split: validation
    path: data/choice_v3/validation.parquet
  - split: test
    path: data/choice_v3/test.parquet
  default: true
- config_name: legacy_replay_v3
  data_files:
  - split: train
    path: data/legacy_replay_v3/train.parquet
  - split: validation
    path: data/legacy_replay_v3/validation.parquet
  - split: test
    path: data/legacy_replay_v3/test.parquet
- config_name: semantic_v3
  data_files:
  - split: train
    path: data/semantic_v3/train.parquet
  - split: validation
    path: data/semantic_v3/validation.parquet
  - split: test
    path: data/semantic_v3/test.parquet
- config_name: tool_context_v3
  data_files:
  - split: train
    path: data/tool_context_v3/train.parquet
  - split: validation
    path: data/tool_context_v3/validation.parquet
  - split: test
    path: data/tool_context_v3/test.parquet
- config_name: tool_environment_holdout_v3
  data_files:
  - split: validation
    path: data/tool_environment_holdout_v3/validation.parquet
  - split: test
    path: data/tool_environment_holdout_v3/test.parquet
- config_name: hermes_actions_v3
  data_files:
  - split: train
    path: data/hermes_actions_v3/train.parquet
  - split: validation
    path: data/hermes_actions_v3/validation.parquet
  - split: test
    path: data/hermes_actions_v3/test.parquet
- config_name: nlupp_hotels_v3
  data_files:
  - split: train
    path: data/nlupp_hotels_v3/train.parquet
  - split: validation
    path: data/nlupp_hotels_v3/validation.parquet
  - split: test
    path: data/nlupp_hotels_v3/test.parquet
- config_name: hint3_native_v3
  data_files:
  - split: train
    path: data/hint3_native_v3/train.parquet
  - split: validation
    path: data/hint3_native_v3/validation.parquet
  - split: test
    path: data/hint3_native_v3/test.parquet
---

# FastBrain-Instruct

**v3 is the only release available on the current `main` branch.** The default configuration is `choice_v3`. The **2026-10-14 v3** data repairs previously documented supervision and evaluation defects. It regenerates semantic candidate pairs, restores full Hermes requests and native call sets, recovers Nemotron schemas, restores legacy validation, and disables unsupported NLU++ Boolean labels. Model training and capability-retention experiments have not been run for this release.

```python
from datasets import load_dataset

ds = load_dataset("TypeScript/FastBrain-Instruct")  # choice_v3
```

## Configurations

| Config | Train | Validation | Test | Purpose |
|---|---:|---:|---:|---|
| `choice_v3` (default) | 113,163 | 9,505 | 21,502 | Integrated repaired choice corpus for the choice head |
| `legacy_replay_v3` | 61,131 | 825 | 7,181 | Legacy coarse routing and restored Hermes single-call choices |
| `semantic_v3` | 26,922 | 7,230 | 12,220 | Regenerated MTOP/TOPv2 candidate-relative match/no-match pairs |
| `tool_context_v3` | 25,110 | 1,450 | 2,101 | Observed Funcdex/Nemotron tool choices with available full schemas |
| `tool_environment_holdout_v3` | — | 534 | 561 | Additional held-out Nemotron policy/catalog groups; evaluation only |
| `hermes_actions_v3` | 868 | 24 | 123 | Native observed call sets; separate auxiliary task |
| `nlupp_hotels_v3` | 765 | 97 | 98 | Native multi-label intents and masked slot values; separate auxiliary task |
| `hint3_native_v3` | 801 | 68 | 1,348 | Native brand-relative intent/OOS labels; separate auxiliary task |

`legacy_replay_v3`, `semantic_v3`, and `tool_context_v3` are exact subsets of `choice_v3`. **Do not concatenate these subsets with the integrated config.** `tool_environment_holdout_v3` overlaps the tool validation/test splits. All three native auxiliary configs have their own schema and are not automatic choice-head training data. Some native Hermes single-call observations also appear in the choice view; `included_in_choice_v3` records this overlap.

Omitting the configuration is equivalent to specifying `"choice_v3"`. All eight configurations belong to the same v3 release; they separate task schemas and evaluation views. Earlier release data, configurations and data cards have been removed from `main`.

## Repairs and evidence

- **Semantic supervision:** all retained pairs are regenerated with three repaired operation descriptions and a conservative conflict graph. All 44 previously confirmed false-no-match cases are repaired or explicitly quarantined. Known context-dependent/ambiguous requests and unsupported persistent call rules are excluded from the choice view. See [case-level outcomes](audit/v3/known_issue_resolution.json) and [the versioned catalog](catalogs/semantic_v3.json). The catalog still contains 68 operations; this is not a 151-operation natural intent dataset.
- **Legacy replay:** restore 825 protected validation rows across all seven legacy sources. Repair documented Git/shell/MASSIVE descriptions, replace a known equivalent pager distractor, and reserve lexical groups against existing holdouts. MASSIVE cooking-label boundaries and other unreviewed labels remain limitations. Missing context is not guessed.
- **Hermes:** all 1,147 original rows are uniquely mapped to the pinned native file, with full requests, system messages, tool schemas and original call sets recovered. Contract failures, documented semantic contradictions and specifically reviewed financial operations are quarantined. Multi-call responses are retained as observed sets in `hermes_actions_v3`; their first call is not made unique gold. Only eligible single-call records with at least two native tools enter the choice view. Source order does not certify parallel independence or a unique execution order.
- **Tool context:** quarantine 31 reviewed inconsistent/ambiguous parents, including contaminated preceding histories. Recover the primary catalog for every published Nemotron source step before filtering: 10,810 native conversations / 23,275 targets. In the final tool view, 23,259 Nemotron rows have full native schemas. Only 10,792 pass both the strict source-prefix check and native argument contract checks needed to expose argument supervision. Other recovered arguments remain explicitly unverified. No execution-correctness or unique-valid-action claim is made.
- **NLU++:** preserve all 960 native texts, intent lists and original slot dictionaries. Disable the generic Boolean export (`head_questions_json={}`, `head_expected_json={}`, `use_for_head_training=false`), because an absent native annotation does not certify a broadly worded predicate false. Mask the 22 reversed stay-date pairs from `slots_for_value_training_json`; retain the original values in `slots_json`.
- **HINT3:** add a separate native auxiliary view of Curekart/SOFMattress with pinned source files and ODbL/DbCL attribution. Source OOS refers to its original brand catalog, not FastBrain's global no-match option. Native test/OOS is never promoted into training. Financing, lexical overlap, empty text and duplicates are handled explicitly; this filtered view is not the canonical HINT3 benchmark.

## Choice adapter and task contract

Use [head_adapter_v3.py](scripts/v3/head_adapter_v3.py) for the current Clef joint-schema format. It assigns deterministic random option IDs per record, preserves the ID-to-name answer mapping, and accounts for the encoder sorting IDs. Merely shuffling the `choices` list does not randomize the encoder's final order. `encode_without_truncation` raises on oversized input instead of letting the upstream encoder crop state.

The documented adapter includes the complete published `query` plus candidate names/descriptions; it does **not** append `tools_json` to state. All 144,170 choice records were encoded with the actual Clef tokenizer/encoder: maximum 11,482 tokens, below 16,384. See [the measured token budget](audit/v3/token_budget.json). Adding full schemas, more head questions or media requires a new budget check.

For observed tool records, `target` is the source-observed action and `acceptable_targets` is a declared positive label list, **not an exhaustive set of valid actions**. Do not turn unobserved alternatives into independently certified `noul=false` or score-zero labels. The supplied adapter exports only the observed choice task. Auxiliary call-set and multi-label records are rejected by that adapter.

## Evaluation and retention

Evaluate both the baseline head and candidate head on the same fixed v3 protocol. Historical results from earlier data are not directly comparable because requests, descriptions, candidates and filtering changed.

The additional Nemotron environment holdout reserves 23 validation and 23 test policy/catalog groups absent from v3 tool training. Existing in-environment validation/test remains in `tool_context_v3` and is identified by provenance. This is exact policy/catalog grouping within this dataset, not proof that a previous model has never seen the public source or semantically similar workflows.

The historical legacy validation is a restored reserved set, not newly unseen data. Historical MASSIVE native-split mixing and unavailable original conversation identities in several legacy sources remain disclosed. Conservative lexical grouping does not establish original dialogue or template ancestry.

Use [fixed evaluation membership](curation/v3/evaluation_membership.jsonl.gz), [evaluation contract](curation/v3/evaluation_protocol.json), and [ablation definitions](curation/v3/ablation_configs.json). Report source/operation-family slices, candidate-count buckets, present/absent results and in-/out-of-environment tool results separately. Paired semantic rows share a parent; uncertainty estimates should group by parent/template/environment as applicable. Predeclare retention tolerances before training; a lower aggregate loss is not a retention gate.

## Provenance, exclusions and limitations

[Release counts and hashes](audit/v3/release_summary.json) · [Validation](audit/v3/validation.json) · [Loading checks](audit/v3/loading_validation.json) · [Repair recipe](docs/curation-v3.md) · [Input evidence](audit/v3/input_manifest.json) · [Quarantine ledger](curation/v3/quarantine.jsonl.gz) · [Historical membership](curation/v3/historical_holdout_membership.json) · [Licenses](LICENSE.md)

The quarantine ledger distinguishes quarantined rows, split reservations, deduplication and records retained in a separate task. It is audit data, not a training config; its row count is not a label-error rate. Native source text and original labels remain available for traceability.

Existing excluded source families/benchmarks remain excluded. Finance filtering is scoped and recorded: it does not establish that every legacy row is finance-free. Ordinary retail price, refund and cash-on-delivery questions are not automatically equated with banking. The semantic 1:1 present/absent construction is an artificial sampling ratio, not a deployment prior. This release does not establish real no-call/clarification policy competence, exhaustive semantic correctness, human gold annotation, or a model-quality gain.

Source acquisition records, source versions and historical input IDs are retained as provenance under `audit/v3` and `curation/v3`. They are not additional dataset releases. The upstream names TOPv2, HINT3 v2, Glaive v2 and Hermes v1 identify their original sources.