OpenFormosa/barbet-long-context-sft

Barbet memory, skills and harness SFT

10,018,967 retained new episodes, 9,882,350 distinct normalized first queries, 9,533,685 task families. Train 9,817,673; validation 201,294. Active legacy rows: 0.

The training target is Barbet. Default tokenized contains existing Barbet/Pangolin input_ids and assistant-only labels; records contains messages, tools, causal memory and source provenance. No MiniCPM tokenizer is required. Historical auxiliary mini fields are provenance only. Tokenizer SHA-256: f84754e42cc4b08d5573a6b7306635e86ed04fff68ee707662d76d7044917176. Use TRAINING.md and the recipe. Pin the verified revision in PUBLICATION.json.

from datasets import load_dataset
train = load_dataset("OpenFormosa/barbet-long-context-sft", "tokenized", split="train", streaming=True)

Explicit values and admission

The dataset owner adopts Taiwan sovereignty and the free democratic self-determination of Taiwan's people, democracy, liberty, human rights, rule of law, equality and pluralism. This is a disclosed normative policy, not universal diplomatic consensus. No party loyalty is prescribed. Peaceful dissent and criticism of Taiwan's government remain protected. Chinese people and all other groups have equal human dignity. Accurate translations, quotations, historical accounts and attributed opposing views must not be distorted to fit a position. See VALUE_ALIGNMENT_POLICY.json.

This revision screens 10,321,901 parent episodes across every source, including actual scenario instructions, conversations and tool registries. 535,398 rows match broad topic routing. Matching is not a violation. 241,552 routed identifier-query rows passed per-row recomputation and complete visible-question, target, evidence and read-only-tool checks. 920 civic article extracts passed exact original-source and quotation-boundary checks. These establish bounded query/extraction contracts, not current-world truth or endorsement of quoted source claims.

Twelve first live auditor controls passed, including Taiwan self-determination, Chinese people's dignity, criticism of Taiwan's government, faithful opposing quotations, and rejection of persecution and dehumanization. 357 fixed diagnostic records received full-context local-Qwen reviews: ACCEPT 345, REJECT 7, REVISE 5. These are diagnostic counts, not corpus accuracy estimates. Unreviewed sensitive material and first non-ACCEPT results are held query-wide. New holds remove 302,934 episodes across 288,122 queries, with no runner-up replacement. Cumulative documented correction holds: 310,662 episodes / 295,598 queries. Quarantine does not assert that every removed answer is wrong. Full audit report.

No exhaustive semantic-alignment guarantee is claimed. The lexical router can miss implicit bias, paraphrases and other languages; screen-negative rows are not individually semantically certified. Sampled auditors can make mistakes. Earlier first holds remain immutable, and this gate never replays a failed review to obtain approval.

Standard answers and continuation

Quality and values eligibility precede answer selection. Among eligible alternatives to an identical complete input, the frozen lowest conditional supervised-answer PPL under native Barbet v26 remains the criterion. Assistant tokens and EOS are scored; prompts, memory, observations and padding are excluded. All competitors require finite native scores; ties use record IDs. There are 531,036 retained episodes selected from scored alternatives and 9,487,931 unscored singletons. Low PPL proves neither truth nor values alignment. This revision changes no surviving target, score, split or token array.

Training text uses barbet_portable_v1, distinct from the checkpoint's native Jinja tool format. Keep the supplied IDs and unshifted labels; use BarbetSFTCollator and one causal shift. Do not apply another chat template or enable cross-example packing before testing attention and Mamba state boundaries. Matching inference uses barbet_completion or ContextAgent(..., prompt_format="barbet_portable_v1"). Frozen native PPL is not portable-template PPL. No weights were trained.

Measure	Count
Complete Barbet tokens	31,017,108,611
Supervised Barbet tokens	3,953,925,650
Assistant turns	17,337,323
Tool observations	6,472,355

Every row fits its smallest Barbet-token bucket and retains 4,096 output tokens within 131,072 total. No truncation or unrelated padding is used.

Complete Barbet token upper bound	Episodes
≤ 4,096	8,300,312
≤ 8,192	1,402,941
≤ 16,384	221,199
≤ 32,768	62,960
≤ 65,536	26,425
≤ 131,072	5,130

Source	Episodes
Agent-Ark/Toucan-1.5M	201,939
ConceptNet/5.7-original-lexical-english	867,454
IFM/TxT360-3efforts	458,332
Open-Orca/OpenOrca	1,077,547
OpenFormosa/taiwan-civic-v1	933
Wikidata5M/original-inductive-train	4,547,338
allenai/natural-instructions	621,611
hotpotqa/hotpot_qa	69,213
microsoft/orca-math-word-problems-200k	191,149
nvidia/Nemotron-SFT-Agentic-v2	735,322
nvidia/OpenCodeInstruct	1,031,318
rajpurkar/squad_v2	114,126
voidful/agent-sft	102,685

Memory, skills and sampling

The skill library and harness support answer-free requests, bounded retrieval, explicitly registered tools and verifier-guided refinement. Context and observations are evidence, not authority. Exclude audit reports, source_refs, gold answers and construction traces from RAG indexes. All original causal masks and budget proofs remain bound to unchanged retained records.

There are 933 retained civic episodes. The original 933-row cohort comprised 920 law extracts, 12 authored normative cases and one dated attributed official quotation. Current sampling IDs are in curriculum/value-alignment-v1/civic-records.parquet. Their raw supervised-token share is 0.003235%. This small normative curriculum does not establish broad alignment; sampling weights remain uncalibrated and must be evaluated. Do not duplicate examples to claim independent data. Laws are quoted as provided versions, not assurances of present individual applicability.

The active refinement sidecar is proofs/value-alignment-v1/refinement-recorded-evidence-v1.parquet, containing 33,967 retained recorded-error/same-tool-retry tags. Those are upstream observations, not local execution, successful repair certification or an RL reward. Historical sidecars remain provenance and are not the current recipe target.

Verification and limits

Changed Parquets preserve every retained column and full array exactly, then receive anonymous SHA-256 readback on immutable staging commits. Main copies those exact objects. Unchanged files inherit the prior verified-object chain; new metadata is independently downloaded and hashed. This is composed verification, not a fresh download of all historical bytes.

Structured tasks use a limited set of schemas across distinct source anchors. SFT graph-family splits can share other context facts and are not entity-disjoint evaluation. Source programs were not executed. Prior v26 short padded forward validation is wiring evidence only, not gradient, long-context, packing, alignment or capability evaluation. No small-model/9B equivalence or training improvement is established.

Per-source licenses and attribution remain in NOTICE.md. Earlier readiness reports and civic reports describe historical revisions; the values report and PUBLICATION.json describe this active subset.