Datasets:

weilan55
/

k3f_regen_data

license: other
license_name: mixed-upstream
license_link: https://huggingface.co/datasets/weilan55/k3f_regen_data#licensing
task_categories:
  - text-generation
language:
  - en
  - zh
tags:
  - speculative-decoding
  - draft-model
  - on-policy
  - multi-turn
size_categories:
  - 1M<n<10M
configs:
  - config_name: v5_train
    data_files: v5_train.jsonl
  - config_name: v6_train
    data_files: v6_train.jsonl
  - config_name: longchat_allturn
    data_files: longchat_allturn.jsonl

K3-Flash regenerated corpora for draft-model training

On-policy responses generated by moonshotai/Kimi-K3-Flash over prompts taken from public datasets, built for training speculative-decoding draft models (EAGLE-3 / DSpark style). The prompts are upstream; every assistant turn here is K3-Flash's own output, which is what makes the data on-policy for a draft that has to predict this target.

No production or service traffic is included. Every row's metadata.source names its upstream public dataset.

Files

file	rows	what it is
`v5_train.jsonl`	293,182	depth-matched sample, deliberately skewed toward bounded asks (69.3% vs 45.4% in the internal baseline)
`v6_train.jsonl`	1,013,914	the full deduplicated pool, no ratio shaping
`longchat_allturn.jsonl`	476,318	all-turn regeneration of 47,651 long conversations: every intermediate assistant turn regenerated against the model's own earlier replies

Schema

v5_train / v6_train — one conversation per row:

{"id": "...", "conversations": [{"role": "user", "content": "..."},
                                {"role": "assistant", "content": "..."}],
 "metadata": {"source": "WildChat-1M", "batch": "E", "n_assistant_turns": 4,
              "depth_band": "4-7", "ask_shape": "bounded",
              "supervised_kind": "text_only", "tail_is_assistant": true}}

longchat_allturn — one generation per row, pretokenized:

{"id": "lc_<conv>_gen2", "primary_id": "lc_<conv>",
 "input_ids": [...], "loss_mask": [0,...,0,1,...,1],
 "conversations": [...prefix..., {"role": "assistant", "content": "..."}],
 "metadata": {"turn_index": 2, "n_turns": 9, "source": "WildChat-1M"}}

loss_mask is 1 exactly over the completion tokens, so input_ids is the true on-policy sequence and the boundary is exact rather than re-derived. Rows of one conversation share a primary_id; turn_index k was generated against a prefix whose earlier assistant turns are the model's own outputs, not the source's.

Upstream prompt sources

source	rows in v6	upstream
OpenCodeInstruct	365,160	`nvidia/OpenCodeInstruct`
kimi-mtp-dataset	172,412	`lightseekorg/kimi-mtp-dataset`
WildChat-1M	148,706	`allenai/WildChat-1M`
aya_dataset	111,545	`CohereLabs/aya_dataset`
ultrachat_200k	99,738	`HuggingFaceH4/ultrachat_200k`
kimi-mtp-llava-mm	97,094	llava subset of kimi-mtp
glaive-function-calling-v2	19,259	`glaiveai/glaive-function-calling-v2`

Known limitations

Multimodal rows carry no pixels. The 97,094 kimi-mtp-llava-mm rows in v6_train use image parts shaped {"type":"image","url":...,"path":...}. Common loaders look for item["image"] or item["image_url"]["url"], so with those this resolves to no image and the row trains as text with a bare <image> placeholder. Affected rows are flagged metadata.mm_image_unreadable: true. The original COCO URLs are in the url field if you want to rehydrate them.

tail_is_assistant: false rows end on a non-assistant message, so a last-turn-only supervision rule has no target on them. Flagged, not removed.

Acceptance is dominated by the supervised turn's kind, not by topic. Measured on these corpora's family: turns that emit a tool call draft at tau ~4.0-4.3 while plain-text turns sit at ~2.0-3.0. Within plain text, a bounded ask (translate, list, extract, fix) drafts +0.4 to +0.65 above an open-ended one; conversation depth is monotonically worse (1 turn ~2.74, 10+ turns ~2.02). CJK text, output length, and markdown structure were each tested and did not explain the gap. These corpora are ~99.9% plain-text-supervised, so they exercise the hard end.

v6_train and longchat_allturn overlap on 29,567 conversations (62% of longchat) — both draw from WildChat/ultrachat. Deduplicate by first-user-turn if you mix them.

Held-out rows matching an internal eval split were removed during the build; this is not a general-purpose decontamination against public benchmarks.

Licensing

Each row's prompt inherits its upstream dataset's license (ODC-BY for WildChat, MIT for ultrachat, and the respective terms for OpenCodeInstruct, aya, glaive and kimi-mtp — check each before redistributing). The assistant turns are generated output from moonshotai/Kimi-K3-Flash and are subject to that model's license. Verify both chains apply to your use case.