README.md

metadata

license: other
tags:
  - sft
  - qwen3.8
  - kimi-k3
  - fable-5
  - nanbeige
  - coding
  - reasoning
  - tool-use
  - 4096-context

Qwen3.8–Kimi–Fable Distillation

A 13,700-row curated SFT mixture of reasoning, coding, instruction, and tool-use traces from three teacher datasets.

Quick start

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="hf://datasets/j0no12/qwen38-kimi-fable-distillation/curated.jsonl",
    split="train",
)

print(dataset[0]["source"])
print(dataset[0]["messages"])

Dataset summary

Property	Value
Rows	13,700
Split	Training collection only
Format	JSON Lines
Context policy	At most 4,096 tokens with the Nanbeige4.2-3B tokenizer
Main domains	Math, code, reasoning, instruction following, tool use
Intended use	SFT for compact agentic models

Composition

source	Rows	Share
qwen3.8-max	10,000	73.0%
fable-5-trace	2,500	18.2%
kimi-k3-trace	1,200	8.8%
Total	13,700	100%

Recorded domain counts are 3,912 math, 2,916 reasoning, 2,650 code, 522 instruction, and 3,700 rows without a domain label.

Schema

Field	Type	Description
messages	list of objects	Chat messages with role and content
source	string	One of the three source labels above
domain	string or null	Domain tag when available
difficulty	source-dependent / optional	Difficulty tag when supplied by the source

Example access pattern:

row = dataset[0]
for message in row["messages"]:
    print(f"{message['role']}: {message['content']}")

Source lineage

Source dataset	Role	Selected rows	Terms recorded by this card
r0b0tlab/qwen3.8-max-distillation-50k	General reasoning, math, code, instruction	10,000	See the source dataset and Alibaba Cloud Model Studio terms
greghavens/fable-5-coding-and-debugging-traces	Coding/debugging trajectories	2,500	CC-BY-4.0
greghavens/kimi-k3-coding-and-debugging-traces	Coding/tool-use trajectories	1,200	CC-BY-4.0

Construction story

The curation report records 61,150 input rows: 44,796 Qwen, 12,448 Fable, and 3,906 Kimi. Quality and length filtering retained 40,689 candidate rows before source caps were applied. The final deterministic mixture contains the 13,700 rows shown above.

Recorded filtering outcomes:

Source	Main recorded exclusions
Qwen	7,413 low-quality scores, 711 over-length contexts, 132 refusals, 6 short assistant responses
Fable	8,196 short assistant responses, 1,380 over-length contexts, 54 refusals
Kimi	2,386 short assistant responses, 174 near-duplicates, 9 refusals

The context check used the Nanbeige4.2-3B tokenizer. User messages range from 5 to 3,027 tokens (median 75), and assistant messages range from 15 to 3,625 tokens (median 259). The raw-content target was approximately 3,800 tokens to leave space for chat-template tokens inside a 4,096-token window.

Quality checks

• Qwen rows required a quality score of at least 7.5; the prior card reports median 9.0 and mean 8.6.
• Refusals and empty/very short completions were filtered.
• Context length was checked with the target tokenizer.
• Kimi processing records a near-duplicate filter; the previous card also described SHA-256 hashing, which should be understood as exact content hashing unless a separate similarity method is supplied.
• Source caps fix the final mixture at 10,000 / 2,500 / 1,200 rows.

Contamination screening

No benchmark-contamination report is included. Hash-based deduplication and source-specific near-duplicate filtering do not prove independence from HumanEval, MBPP, GSM8K, MATH, or other public evaluations. Run task-specific decontamination before reporting benchmark results from models trained on this collection.

Known limitations

• All responses are teacher-generated and may contain incorrect reasoning, insecure code, fabricated tool results, or stylistic artifacts.
• There is no held-out validation or test split.
• The 4,096-token guarantee is tokenizer-specific and may not hold under another tokenizer or chat template.
• Domain labels are missing for 3,700 rows, and optional metadata is source-dependent.
• The mixture intentionally oversamples Qwen relative to the two trajectory sources.
• Component licensing is mixed; no single permissive license is asserted for every row.

Files and integrity

File	Size	Integrity identifier
curated.jsonl	64,322,634 bytes	SHA-256 78d2d52229b49e057fad41cf3ec66a60e5be7914997b55938cb3ba863365bb8a
report.json	1,303 bytes	Git blob 095b1324ac986777402e3cb58a1fbfa792a67fd5
Repository revision documented by this card update	—	95156b7351b5a7689b5aa6eee037337d3ee5293f

Citation

@misc{j0no12_qwen38_kimi_fable_2026,
  author       = {j0no12},
  title        = {Qwen3.8--Kimi--Fable Distillation},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/j0no12/qwen38-kimi-fable-distillation}}
}

Please also cite the three source datasets.

License

Mixed / other. Fable and Kimi components are recorded as CC-BY-4.0; Qwen-derived traces remain subject to their source dataset and Alibaba Cloud Model Studio terms. This card does not grant rights beyond those upstream terms.