Qwen3.8-Flash-Next-DSpine-generations

# Qwen3.8-Flash-Next DSpine research generations An attributed, conservative subset of on-policy target generations prepared for speculative-draft research. **73,157 records, 176,812,512 response tokens**, 18 parquet shards. The full retained generation inventory was 100,318 records / 228,480,448 response tokens. Five sources were excluded in full after a per-source license audit and owner review. **There is no blanket Apache-2.0 license for these prompts.** Each row has a `source_license` and provenance. The prompt sources retain their individual licenses and attribution obligations; see [LICENSES.md](LICENSES.md) and [sources.json](sources.json). New model outputs do not erase prompt rights or relicense copied underlying content. This collection imposes no additional blanket license over third-party material. The separate draft-head/code license is not a grant covering this dataset. ## Included sources and attribution Original final answers were removed; queries and earlier conversation/tool context were retained. Responses are new target-model continuations. Source publishers and contributors are credited below; pinned revision/file/row details are preserved in each row. Counts describe this exported subset, not upstream dataset sizes. | Source dataset | Attribution | Prompt license | Records | Response tokens | |---|---|---|---:|---:| | [AI-MO/NuminaMath-CoT](https://huggingface.co/datasets/AI-MO/NuminaMath-CoT) | AI-MO and contributors | apache-2.0 | 4,134 | 12,277,826 | | [CohereForAI/aya_dataset](https://huggingface.co/datasets/CohereForAI/aya_dataset) | CohereForAI and contributors | apache-2.0 | 5,016 | 9,238,437 | | [HuggingFaceH4/ultrachat_200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) | HuggingFaceH4 and contributors | mit | 9,957 | 27,393,024 | | [NousResearch/hermes-function-calling-v1](https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1) | NousResearch and contributors | apache-2.0 | 1,161 | 620,846 | | [OpenAssistant/oasst2](https://huggingface.co/datasets/OpenAssistant/oasst2) | OpenAssistant and contributors | apache-2.0 | 2,644 | 6,472,747 | | [THUDM/LongCite-45k](https://huggingface.co/datasets/THUDM/LongCite-45k) | THUDM and contributors | apache-2.0 | 2,040 | 5,723,439 | | [Team-ACE/ToolACE](https://huggingface.co/datasets/Team-ACE/ToolACE) | Team-ACE and contributors | apache-2.0 | 8,467 | 2,550,331 | | [allenai/WildChat-1M](https://huggingface.co/datasets/allenai/WildChat-1M) | allenai and contributors | odc-by | 9,538 | 27,009,296 | | [bigcode/self-oss-instruct-sc2-exec-filter-50k](https://huggingface.co/datasets/bigcode/self-oss-instruct-sc2-exec-filter-50k) | bigcode and contributors | odc-by | 4,686 | 12,928,612 | | [glaiveai/glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2) | glaiveai and contributors | apache-2.0 | 4,752 | 589,494 | | [gretelai/synthetic_text_to_sql](https://huggingface.co/datasets/gretelai/synthetic_text_to_sql) | gretelai and contributors | apache-2.0 | 3,832 | 2,598,768 | | [ise-uiuc/Magicoder-OSS-Instruct-75K](https://huggingface.co/datasets/ise-uiuc/Magicoder-OSS-Instruct-75K) | ise-uiuc and contributors | mit | 4,631 | 21,616,813 | | [meta-math/MetaMathQA](https://huggingface.co/datasets/meta-math/MetaMathQA) | meta-math and contributors | mit | 3,764 | 2,077,831 | | [nvidia/Nemotron-Agentic-v1](https://huggingface.co/datasets/nvidia/Nemotron-Agentic-v1) | nvidia and contributors | cc-by-4.0 | 1 | 191 | | [nvidia/OpenCodeInstruct](https://huggingface.co/datasets/nvidia/OpenCodeInstruct) | nvidia and contributors | cc-by-4.0 | 4,547 | 11,820,246 | | [nvidia/OpenMathReasoning](https://huggingface.co/datasets/nvidia/OpenMathReasoning) | nvidia and contributors | cc-by-4.0 | 3,987 | 33,894,611 | WildChat uses ODC-BY, retroactively replacing its earlier ImpACT license. StarCoder2 Self-Alignment also uses ODC-BY. That license covers database rights, not independently copyrighted content. Do not assume it relicenses code copied inside a prompt. The export does not include unused raw seed-code fields. WildChat IP hashes, headers, locations, timestamps and source user IDs were not part of prompt extraction. Redacted/toxic conversations and known credential patterns were filtered in the original prompt build. OpenMathReasoning contributes only its selected `additional_problems` split. The selected Terminal-Corpus and neulab components have no generated records in this retained inventory, so neither is included in this dataset. ## Sources excluded after review | Source | Removed records | Removed response tokens | Reason | |---|---:|---:|---| | SWE-bench/SWE-smith-trajectories | 8,124 | 2,406,806 | MIT dataset card does not settle licenses/attribution of embedded repository code in tool context. | | nebius/SWE-agent-trajectories | 4,168 | 1,664,516 | Card requires repository-specific licenses and notes Llama-output obligations. | | m-a-p/CodeFeedback-Filtered-Instruction | 4,632 | 29,199,546 | Named upstream branches include NC/ND and NC/SA terms despite the aggregator's Apache card. | | mlabonne/open-perfectblend | 7,226 | 14,955,586 | Mixed upstream chain, including CodeAlpaca-derived material with NC data terms; the whole source was omitted. | | nvidia/Llama-Nemotron-Post-Training-Dataset | 2,950 | 3,220,680 | Per-sample CC-BY/ODC-BY/CC-BY-SA and possible Llama terms were not retained as row licenses in extracted prompts. | These exclusions remove 27,100 records and 51,447,134 response tokens. Later row-level upstream/repository-license joins could recover eligible portions; none was performed for this conservative version. LMSYS-Chat-1M, unknown-license, noncommercial and no-derivatives candidates are not admitted. The differently licensed Arena preference subset is not LMSYS-Chat-1M, but it is still omitted here as part of the whole Open-PerfectBlend exclusion. The complete review is in [dataset-license-audit.md](dataset-license-audit.md). ## Generation model, tokenizer and settings Target: [RedHatAI/Qwen3.8-Flash-Next-NVFP4](https://huggingface.co/RedHatAI/Qwen3.8-Flash-Next-NVFP4), revision `c8f2fb1b9869f686b214782036123b10ff96d14a`. The tokenizer and chat template are from the same snapshot; tokenizer file hashes are in `manifest.json`. Temperature **1.0**, top-k **20**, top-p **0.95**. Exact prompt token IDs were sent to the generation server. Outputs were decoded using the pinned tokenizer, `skip_special_tokens=False`, `clean_up_tokenization_spaces=False`. Output IDs are authoritative; reasoning/tool/end markers remain in the response text. Responses are outputs of Qwen3.8-Flash-Next (through the RedHatAI NVFP4 quantization), which is distributed under the Qwen Community License 1.0; a copy is in `licenses/Qwen-Community-License-1.0.txt`. Its conditions, including those for Model-as-a-Service and AI Work Assistant businesses, may apply to uses of these outputs. The actual thinking flags are **67,143 true and 6,014 false**, with no unknown flags. Reasoning effort: **36,271 xhigh, 22,026 medium, 14,860 low**. Original per-prompt `chat_template_kwargs` and tools are retained. Do not describe every record as thinking enabled. Finish reasons: 67,733 stop and 5,424 length-capped. Serving numerical differences across generation platforms remain possible; the nominal T/k/p settings alone do not imply bit-identical output laws. ## Schema and stable validation split | Column | Meaning | |---|---| | `id` | Stable canonical prompt ID. | | `source`, `source_license`, `domain` | Attributed source, prompt license and domain. | | `prompt_messages` | Role/content text messages; content is nullable where tool calls carry the message. | | `prompt_messages_full_json`, `tools` | Original full message dictionaries/tool schemas as JSON, preserving tool calls and IDs. | | `response`, `output_token_ids` | Decoded continuation and original generated token IDs. | | `finish_reason`, `sampling` | Original completion status and T/k/p. | | `enable_thinking`, `reasoning_effort`, `chat_template_kwargs` | Actual prompt-specific thinking/template metadata. | | `provenance` | Source tag/revision/file/row/upstream source/URL, generation model/revision/platform family/part. | | `is_validation` | The trainer's existing deterministic 1% hash split. | There are **72,452 training rows and 705 validation rows**. The rule is the first 16 hex digits of SHA-256(id), divided by 2**64, less than 0.01. This is distinct from the separate 2,000 prompt-family holdouts used to produce evaluation caches. The parquet container is loaded as one HF `train` split; **filter the flag before training**, rather than training on every loaded row: ```python from datasets import load_dataset data = load_dataset("ProCreations/Qwen3.8-Flash-Next-DSpine-generations", split="train") train = data.filter(lambda row: not row["is_validation"]) validation = data.filter(lambda row: row["is_validation"]) ``` ## Filtering, intended use and limitations Use for research on target imitation, speculative draft heads and sampling. Do not treat these synthetic continuations as verified answers. They can contain mistakes, harmful content and copied prompt text. This is a generated subset of a deliberately mixed prompt pool, not a representative sample of all use cases. The export honors the two-ID `EXCLUDE_IDS.json` and exact/near prompt dedup. Prompt decontamination used normalized 13-grams against GSM8K test, MATH-500, HumanEval, MBPP test, LiveCodeBench, MT-Bench, Arena-Hard and local serving prompts; a supplemental tool-argument audit produced the two excluded IDs. Neither ID appears in this export. These procedures reduce known overlap and are not proof of zero contamination. The exclusion file/hash and audit are included. The full-row validation checks unique IDs, allowed sources/licenses, source provenance, target revision, sampling, token totals, deterministic split and shard hashes; sample decoding is checked against the local tokenizer. 61 rows were removed by a credential-pattern scan. No target features, probability caches, optimizer states or target weights are included. Files and checksums are listed in `manifest.json`; validation and credential-scan scope are recorded in `validation-receipt.json`. Credit the source dataset authors above, Qwen/Red Hat for the target, and the [DSpine](https://arxiv.org/abs/2609.36173) and [DFlash](https://github.com/z-lab/dflash) authors. This independent collection is not endorsed by those projects.