ProCreations/Nanbeige4.3-33B-Distill

Nanbeige4.3-33B Distill

Staged review edition. Upload requires a separate approval.

Attributed source conversations and Nanbeige4.3-33B continuations for speculative drafter training and quant calibration. This is a licence, hygiene and test-filtered edition of the original generation pools, not the complete original training corpus. Default config: drafter_v2.

Configs

Config	Rows	Prompt tokens	Response tokens	Total tokens	Use
drafter_v1	26,690	124,618,730	54,554,973	179,173,703	Drafter training
drafter_v2	33,856	107,778,758	86,779,983	194,558,741	Drafter training
calibration	1,062	1,330,474	1,932,226	3,262,700	Quant calibration
eval	3,892	9,608,445	10,431,402	20,039,847	DO NOT TRAIN; test only

eval has only a test split and every row has do_not_train=true. Its banks field distinguishes original eval-v1, all allowed generated heldout prompts and internal-validation/public-test material removed from drafter training. The same native row shared by reference views appears only once, with multiple bank tags.

The public calibration and eval configs are not byte-identical to the private calibration/eval used for the published quant measurements. Rights and hygiene filtering changes their rows and domain balance. Do not reproduce the published scores or claim the original 35% agentic calibration mix from this filtered edition. The eval config is broader than the original 272-row eval-v1 bank.

Final domain counts

Tokens use the pinned Nanbeige tokenizer. Prompt tokens include the complete rendered upstream history; response tokens count only the newly generated continuation.

Config	Domain	Rows	Prompt tokens	Response tokens	Total tokens
drafter_v1	agentic	13,358	119,748,788	21,858,628	141,607,416
drafter_v1	chat	3,098	1,378,433	8,619,472	9,997,905
drafter_v1	chinese	949	359,541	2,328,375	2,687,916
drafter_v1	code	4,417	992,677	13,798,647	14,791,324
drafter_v1	math	1,598	166,640	2,110,470	2,277,110
drafter_v1	multilingual	1,325	313,384	4,531,792	4,845,176
drafter_v1	stem	4	2,400	11,408	13,808
drafter_v1	structured	1,941	1,656,867	1,296,181	2,953,048
drafter_v2	agentic	10,825	99,426,111	19,763,594	119,189,705
drafter_v2	chat	6,901	3,235,148	20,762,820	23,997,968
drafter_v2	chinese	2,984	704,800	14,106,105	14,810,905
drafter_v2	code	3,300	777,322	12,487,504	13,264,826
drafter_v2	math	3,337	336,483	5,326,726	5,663,209
drafter_v2	multilingual	3,167	696,997	12,413,935	13,110,932
drafter_v2	stem	21	6,517	51,975	58,492
drafter_v2	structured	3,321	2,595,380	1,867,324	4,462,704
calibration	agentic	175	960,329	108,871	1,069,200
calibration	chat	123	76,478	397,120	473,598
calibration	chinese	73	39,057	266,570	305,627
calibration	code	286	60,971	521,708	582,679
calibration	math	270	27,439	292,785	320,224
calibration	multilingual	51	12,947	189,547	202,494
calibration	stem	30	19,067	102,416	121,483
calibration	structured	54	134,186	53,209	187,395
eval	agentic	993	8,535,211	1,797,346	10,332,557
eval	chat	683	322,270	2,469,002	2,791,272
eval	chinese	212	71,655	577,609	649,264
eval	code	682	158,268	2,911,031	3,069,299
eval	math	314	31,435	577,644	609,079
eval	multilingual	291	81,786	1,415,260	1,497,046
eval	stem	24	4,988	140,784	145,772
eval	structured	693	402,832	542,726	945,558

Final source counts

Config	Source	Rows	Prompt tokens	Response tokens
drafter_v1	AI-MO/NuminaMath-CoT	807	89,287	1,759,857
drafter_v1	CohereForAI/aya_dataset	874	154,218	2,957,763
drafter_v1	HuggingFaceH4/ultrachat_200k	1,913	1,065,800	5,224,984
drafter_v1	NousResearch/hermes-function-calling-v1	488	435,326	371,438
drafter_v1	OpenAssistant/oasst2	807	53,909	2,223,385
drafter_v1	allenai/WildChat-1M	1,908	842,519	5,593,777
drafter_v1	bigcode/self-oss-instruct-sc2-exec-filter-50k	1,550	226,930	3,899,247
drafter_v1	glaiveai/glaive-function-calling-v2	493	160,989	68,960
drafter_v1	gretelai/synthetic_text_to_sql	659	106,128	506,599
drafter_v1	ise-uiuc/Magicoder-OSS-Instruct-75K	1,299	412,089	5,989,208
drafter_v1	meta-math/MetaMathQA	813	72,070	392,742
drafter_v1	nvidia/Nemotron-Agentic-v1	267	933,796	205,782
drafter_v1	nvidia/Nemotron-Terminal-Corpus	13,358	119,748,788	21,858,628
drafter_v1	nvidia/OpenCodeInstruct	1,454	316,881	3,502,603
drafter_v2	AI-MO/NuminaMath-CoT	1,678	181,737	4,336,995
drafter_v2	CohereForAI/aya_dataset	2,321	424,336	9,652,567
drafter_v2	HuggingFaceH4/ultrachat_200k	4,418	2,480,985	12,845,070
drafter_v2	NousResearch/hermes-function-calling-v1	282	155,728	276,503
drafter_v2	OpenAssistant/oasst2	1,173	76,316	3,274,271
drafter_v2	Team-ACE/ToolACE	1,104	762,778	524,904
drafter_v2	allenai/WildChat-1M	5,245	1,697,673	21,949,073
drafter_v2	bigcode/self-oss-instruct-sc2-exec-filter-50k	1,058	160,851	3,336,119
drafter_v2	glaiveai/glaive-function-calling-v2	780	261,621	124,023
drafter_v2	gretelai/synthetic_text_to_sql	781	130,398	688,557
drafter_v2	ise-uiuc/Magicoder-OSS-Instruct-75K	1,108	352,799	5,739,767
drafter_v2	meta-math/MetaMathQA	1,661	152,230	963,139
drafter_v2	nvidia/Nemotron-Agentic-v1	371	1,283,129	244,734
drafter_v2	nvidia/Nemotron-Terminal-Corpus	10,825	99,426,111	19,763,594
drafter_v2	nvidia/OpenCodeInstruct	1,051	232,066	3,060,667
calibration	AI-MO/NuminaMath-CoT	128	13,839	198,292
calibration	CohereForAI/aya_dataset	21	5,130	95,655
calibration	HuggingFaceH4/ultrachat_200k	104	67,683	306,194
calibration	NousResearch/hermes-function-calling-v1	5	4,003	7,305
calibration	OpenAssistant/oasst2	27	1,723	95,675
calibration	Team-ACE/ToolACE	2	2,349	1,911
calibration	allenai/WildChat-1M	132	74,839	468,474
calibration	bigcode/self-oss-instruct-sc2-exec-filter-50k	99	14,838	154,697
calibration	gretelai/synthetic_text_to_sql	8	1,553	23,853
calibration	ise-uiuc/Magicoder-OSS-Instruct-75K	69	21,552	166,881
calibration	meta-math/MetaMathQA	142	13,600	94,493
calibration	nvidia/Nemotron-Agentic-v1	39	126,281	20,140
calibration	nvidia/Nemotron-Terminal-Corpus	175	960,329	108,871
calibration	nvidia/OpenCodeInstruct	111	22,755	189,785
eval	AI-MO/NuminaMath-CoT	154	16,926	509,074
eval	CohereForAI/aya_dataset	174	35,497	811,273
eval	HuggingFaceH4/ultrachat_200k	430	240,882	1,548,681
eval	NousResearch/hermes-function-calling-v1	339	240,476	221,469
eval	OpenAssistant/oasst2	211	14,984	930,188
eval	allenai/WildChat-1M	450	205,432	1,652,239
eval	bigcode/self-oss-instruct-sc2-exec-filter-50k	231	35,480	699,809
eval	glaiveai/glaive-function-calling-v2	272	88,156	41,857
eval	gretelai/synthetic_text_to_sql	28	4,515	32,616
eval	ise-uiuc/Magicoder-OSS-Instruct-75K	222	70,961	1,493,320
eval	meta-math/MetaMathQA	160	14,326	68,350
eval	nvidia/Nemotron-Agentic-v1	18	57,627	12,621
eval	nvidia/Nemotron-Terminal-Corpus	993	8,535,211	1,797,346
eval	nvidia/OpenCodeInstruct	210	47,972	612,559

Generation and provenance

Teacher: Nanbeige/Nanbeige4.3-33B at c1093c501cc6460427c0ac634db7223d0450c0ee, BF16 weights. V1 KV is BF16 or FP8 E4M3, recorded per row; v2 production KV is FP8 E4M3. Calibration/eval reference generation uses BF16 KV; eval rows moved from v1 retain their actual per-row KV. Optional FP8-weight pilot rows are not included.

The actual Nanbeige chat template uses XML tool calls, the original thinking-on/off flag and preserved reasoning for multi-turn tool histories. Sampling is temperature 1.0 for agentic/tool tasks and 0.6 otherwise, top_p 0.95. Caps were 16,384 new tokens in early v1/reference runs and 8,192 in later bulk/v2 runs. Per-row sampler metadata is authoritative. Recorded seeds are preserved; unknown historical seeds are null with seed_recorded=false, never invented.

Native prompt_ids and response_ids are the exact stored IDs seen and produced by the teacher. A generated EOS is retained. finish_reason, is_truncated and response_complete distinguish EOS-complete rows from capped prefixes. No artificial EOS is appended.

Multi-turn histories are upstream agent transcripts. Only the final appended continuation is generated by Nanbeige; this is not a claim that historical assistant/tool actions were generated or executed by Nanbeige. Tools were not executed during continuation generation.

Format and use

Each Parquet row includes messages with role/content and separate reasoning_content; parsed tool calls with JSON argument strings and retained raw XML; linked tool-result messages; tool schemas; native ID arrays and true lengths; source dataset/revision/file/row/licence; content/task family hashes; teacher revision, precision, sampler and nullable seed; finish/truncation flags and parse/fidelity metadata. Calibration/reference rows also carry exact original segment and decision-position metadata where available.

Tool-call/result IDs reconstructed from historical order are explicitly marked inferred. Incomplete or ambiguous tool prefixes remain in readable content with parse flags. XML content is preserved as well as its structured view. Thinking can be incomplete on capped responses. Native IDs are authoritative: decoding and re-encoding may change byte-fallback segmentation, so do not replace the native IDs with re-tokenized display text. The per-row decoded_reencode_matches_native flag reports that check.

Use only drafter train configs for training. Load calibration intentionally for quant calibration; do not concatenate eval/test into training. When selecting further train/validation splits, split by source/task family aliases, not row IDs. Histories and sibling samples may share task families across v1/v2. The release is not a universally decontaminated benchmark.

Licence, filters and limitations

See ATTRIBUTION.md, LICENSES/README.md, SOURCES.json and the per-row licence. Mixed Apache-2.0, MIT, CC BY 4.0 and ODC-BY terms remain applicable. No included source has a declared non-commercial or share-alike dataset licence. Database licences do not settle all independent rights in source content.

Chinese and math coverage is reduced by rights filtering. OpenMathReasoning and LongCite are held; Numina AoPS-derived tags are excluded. Other Chinese rows remain from approved sources, including WildChat. OpenHands is held; included real agentic histories come from the approved Terminal component. This public edition has different domain proportions from the generation targets and original private corpus.

Upstream WildChat redacted/toxic rows were skipped at ingestion. The publication filter drops entire rows with any PII/credential/private-URL/project-infrastructure detector hit, including example-style email strings. User/IP/country/timestamp/hashed-user/request-header metadata is never copied. The structured view and serialized Parquet are checked again. Dropping whole rows also removes their native token arrays; no unredacted hidden payload is retained in a redacted text column.

Known heldout and evaluation-family registries and 6,982 normalized public-test prompt hashes (GSM8K, HumanEval, MBPP and MATH) exclude matching families from train/calibration. Such rows can appear only in the labelled eval config. Exact native payloads and duplicate IDs are checked; reference views are merged. This does not establish universal semantic benchmark decontamination.

Regex hits are candidates, not confirmed active secrets. No credential was tested. No perfect PII-free claim is made: names, addresses, semantic identifying details and unknown secrets may escape these detectors. Upstream and generated content may be inaccurate, unsafe or copyrighted. Tool calls are data, not instructions to execute.

No standalone prompt-bank or upstream-agent-history config, optional FP8-weight pilot, target features, hidden-state captures, caches, weights or training states are included.

Integrity and review

MANIFEST.json lists every release payload with byte size and SHA256. STAGED_READY.json seals the manifest and core documents. COUNTS.json contains final per-config/domain/source counts, finish reasons, parser statistics and post-format whole-row hygiene drops. Five deterministic uniformly ranked random decoded rows per config are supplied separately for review, with full histories and no truncation.

Config	EOS-complete rows	Length-capped rows
drafter_v1	24,550	2,140
drafter_v2	29,662	4,194
calibration	1,062	0
eval	3,614	278

Final structuring/combined-decoding hygiene drops: 553. Post-serialization detector hits in retained row content and metadata: 0. Exact duplicate public IDs: 0.