# Curation record: 2026-10-14 v2

This release was built from the fixed inputs listed in [source_manifest.json](../audit/source_manifest.json). Per-row provenance and the catalogs record the accepted source coordinates and adaptation. Counts refer to this frozen release, not to the full upstream datasets. The sampling seed is `20261014`. Python and library versions are recorded in the manifest.

## Included sources and task boundaries

MTOP English and TOPv2 are one semantic source family. Funcdex supplies synthetic multi-turn function-calling conversations. Nemotron tool-selection steps come from `MaziyarPanahi/AgentToolDecisions-180K` at `22b91136648789a2881183dfc2f77d756659444c`; all 37,233 inspected tool-selection records trace to `nvidia/Nemotron-SFT-Agentic-v2`, revision `7c804833427f633ccd53b582dbf02525fd680f78`, config `default`, partition `interactive_agent`. No When2Call portion or separate Nemotron `tool_calling` partition is included.

NLU++ uses original English hotel files from PolyAI's repository at `57ec275d8078af65b7731c2a98be812d844a6d6b`. Banking data and Funcdex toolkits containing Stripe are excluded, following the original corpus's finance scope. Bitext and HWU64 remain candidates for separate template/family audits and are not part of this release.

## Semantic adaptation

1. Validate bracketed source parse trees and MTOP's declared root intent; retain original query text and source coordinates. Do not promote source evaluation/test rows to training.
2. Apply the published [ontology mapping](../curation/semantic_ontology_review.json). Merge aliases, retain reviewed descriptions and identify overlapping or parent/child operations. Single-choice candidates require a single intent node and a usable reviewed label. Native `UNSUPPORTED_*` labels are source-relative and never become global no-match labels automatically.
3. Exclude benchmark lexical matches and exact legacy query overlaps. Resolve conflicting mapped labels conservatively. Normalize with Unicode NFKC, case folding and word-token extraction. Replace leaf slot spans and numeric tokens to derive template keys.
4. Assign duplicates/templates by source split priority: test, then validation, then train. Remove other-split instances; deduplicate normalized queries. These operations use text and provenance for isolation, not held-out model performance.
5. Apply the [source screening rules](../curation/semantic_query_screening_rules.json) and [additional training-sample review rules](../curation/quality_overrides.json). The review found mixed feedback/volume requests, repeat-versus-previous ambiguity and contextless call/navigation operations. It also clarified that education-time descriptions cover duration. Rules are conservative rejection rules, not relabeling rules. They do not equate missing execution slots with unsupported intent.
6. Require at least 20 independent training requests after isolation. Cap each class at 400 train and 150 validation/test requests, with deterministic seeded ordering. The resulting main catalog has 68 labels. Complex, ambiguous, native unsupported and low-support records go to `semantic_original_labels` with their original labels and exclusion reasons.
7. For every accepted request, construct one positive and one candidate-relative no-match row in the same split. Include non-equivalent neighbors and random other catalog entries. Conflict exclusions are symmetric for all listed conflict types. Replace the removed target with another non-equivalent candidate, keep the same candidate count and include the same no-match option on both sides. Shuffle candidate positions. The catalog hash includes descriptions and excluded-neighbor sets.

This yields 13,488 independent training requests and 26,976 paired training rows. Candidate catalogs are an adaptation of reviewed source labels; the review does not certify every utterance's sole valid interpretation. All pair variants must remain grouped by `parent_id` when resplitting.

## Tool adaptation

**Funcdex.** Keep native test conversations in test and assign approximately 10% of eligible native train conversations to validation by a deterministic hash of the normalized initial request. Group by that initial request. Reject whole conversations for duplicate/unknown tool names, invalid or multiple calls, malformed arguments, schema failures, benchmark overlaps, excluded finance toolkits and known synthetic/unsupported-result scaffolding. Preserve the source system message and every preceding message before each selected assistant call. The current answer is not included in the input. Preserve full tool schemas and original target arguments in structured metadata.

Targeted inspection identified 13 conversations rejected by the additional review rules, including a drive-creation trajectory lacking policy prerequisites and a trajectory claiming an assumed spreadsheet result. This is not a random-sample estimate of the source's error rate. The remaining labels record observed calls, not certified unique actions. No-call, clarification, confirmation and completion turns are not converted to negative tool-choice examples.

**Nemotron.** Map all 36,048 original Nemotron records back to source-derived rows by query, target and candidate reconstruction. The original train/test split shares 2,687 raw parent conversations; previous augmented train versus historical validation/test shares 108 parents. Block the union of 3,286 protected parents. Replace all 32,500 cropped Nemotron training rows in the v2 configuration with eligible source-derived prefixes. Preserve frozen legacy test records for continuity.

The intermediate source kept at most 12 conversation messages and truncated some message objects. Its separate policy field can repair truncated system messages. Restore that policy and discard steps with known missing earlier history, non-system truncation or unavailable policy. Apply benchmark and finance exclusions to entire raw conversations. The resulting tool view has 22,231 Nemotron training steps; 5,574 kept steps across all splits required system-truncation repair. Full original function schemas were not recovered. [Parent blocklist](../curation/nemotron_reserved_parents.json), [canonical lineage](../curation/nemotron_canonical_lineage.jsonl.gz), [context audit](../audit/nemotron_context_repair_summary.json).

## Legacy replay and NLU++

Retain 63,223 original non-Nemotron training records after frozen historical holdout-ID/query exclusion, normalized duplicate removal and the recorded lexical benchmark checks. Preserve core fields. Legacy sources other than mapped Nemotron do not all expose parent dialogue IDs, so query-level isolation does not prove conversation/template independence for those sources. [Legacy holdout identifiers](../curation/legacy_holdout_ids.json) preserve existing public-record membership for future exclusions; they are not a newly independent evaluation set.

NLU++ hotel folds 0–15 are train, 16–17 validation and 18–19 test. Remove exact conflicts and cross-split query/slot-template overlaps. Preserve every annotated intent module and slot; nonempty-intent rows get optional 40-module boolean supervision. Empty-intent records remain context fragments with no generated head labels. The resulting train/validation/test intent-eligible counts are 576/71/70.

## Exclusion and validation scope

The [source-family policy](../curation/excluded_source_families.json) retains all 35 ToolRet source exclusions, plus the existing benchmark exclusions and their known derivatives. Source-family exclusion is applied before lexical checks; rewriting a disallowed source does not make it eligible.

The [lexical index manifest](../audit/exclusion_manifest.json) records the exact cached inputs: selected states for ten benchmarks, all 35 ToolRet query-source files, CLINC data, cached BANKING77, BFCL v3 requests and cached API-Bank user turns. Matching uses normalized exact text or a shared ten-word sequence. Coverage is limited to those files; in particular, selected-state coverage is not every split of every benchmark. Slot-template checks apply to new semantic and NLU++ families. No model-based paraphrase decontamination claim is made. Exclusion hashes are included as audit controls, not training data.

[Validation](../audit/validation.json) checks every released record's basic contract and hashes, source coordinates, original-field preservation, view identity, positive/negative pair construction, candidate conflicts, tool prefix/label alignment, recorded context-loss flags, protected parents and NLU++ multi-label preservation. Cross-split ID, parent, normalized-query and applicable template contracts are checked. Mechanical validation does not replace semantic adjudication, and no model training or retention-gate experiment is part of this dataset release.