Datasetpublic

DaoCloud/Qwen3.8-27B-Drafter-SFT

Qwen3.8-27B Drafter SFT Corpus Supervised fine-tuning data released for training speculative drafters for Qwen/Qwen3.8-27B. All completions were generated with Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. The dataset contains 367,535 source conversations and 450,401 train rows, totaling 1,953,218,671 tokens after filtering and evaluation decontamination. Rows contain Qwen3.8-27B-tokenized prompts and target-generated completions, together with loss… See the full description on the dataset page: https://huggingface.co/datasets/DaoCloud/Qwen3.8-27B-Drafter-SFT.

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

1likes844downloads

Dataset Card

Qwen3.8-27B Drafter SFT Corpus

Supervised fine-tuning data released for training speculative drafters for `Qwen/Qwen3.8-27B`. All completions were generated with `Qwen/Qwen3.8-27B` at revision `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`.

The dataset contains 367,535 source conversations and 450,401 train rows, totaling 1,953,218,671 tokens after filtering and evaluation decontamination. Rows contain Qwen3.8-27B-tokenized prompts and target-generated completions, together with loss masks selecting the supervised terminal assistant span.

Sources

Source	Conversations	Rows	Prompt tokens	Completion tokens	Upstream terms
`mlabonne/open-perfectblend`	209,120	291,986	533,670,473	761,253,172	Apache-2.0
`nvidia/OpenCodeInstruct`	59,827	59,827	16,675,325	313,827,296	CC-BY-4.0
`CohereLabs/aya_dataset`	58,835	58,835	7,608,239	81,935,536	Apache-2.0
`nvidia/Nemotron-RL-Instruction-Following-Structured-Outputs-v2`	19,993	19,993	64,493,674	47,542,725	CC-BY-4.0
`nvidia/Nemotron-SFT-Instruction-Following-Chat-v3`	19,760	19,760	71,257,395	54,954,836	CC-BY-4.0 / ODC-By; see source-level notices
Total	367,535	450,401	693,705,106	1,259,513,565	Mixed

Schema

Field	Description
`input_ids`	Qwen3.8-27B-tokenized prompt and target-generated completion
`loss_mask`	Terminal assistant-token supervision mask
`seq_len`	Sequence length
`source`	Normalized source identifier
`row_id`	Unique train-turn identifier
`primary_id`	Source conversation identifier
`continue_final_message`	Whether tokenization continues the final source message
`reasoning_effort`	`low`, `medium`, or `xhigh`

Completions were generated by the target model in thinking mode with reasoning_effort set to low, medium, or xhigh. The approximate mixture is 40% / 30% / 30%.

Tokenization

All rows are tokenized with the tokenizer and chat template for `Qwen/Qwen3.8-27B`. Target hidden states are not included.

Prompts

The prompts config carries the untokenized input side of all 367,535 conversations, joined to the default config on primary_id. messages holds system and user turns only — assistant turns are target-model generations and ship tokenized in the default config as input_ids[loss_mask]. tools, upstream_source, and language are populated where the source supplies them.

Decontamination

The corpus was filtered for degenerate repetition and decontaminated against the 5,238-prompt evaluation suite used for drafter evaluation.

Files

Training data is released as Parquet shards:

text

data/train-*.parquet

The canonical training representation is input_ids together with loss_mask.

License

This release combines datasets with different licenses and usage terms and is therefore marked license: other.

Prompts derive from the upstream datasets listed above and remain subject to their terms. Completions were generated by `Qwen/Qwen3.8-27B`, released under Apache-2.0, and are redistributed under those terms.

See `THIRD_PARTY_LICENSES.md` and the corresponding upstream dataset cards for source-level licensing and attribution information.