Datasetpublic

ArkhAngelLifeJiggy/Qwen3.8-27B-Distillation-40K

Qwen3.8-27B-Distillation (40K Traces) Qwen3.8-27B-Distillation is a dataset containing 40,000 reasoning traces distilled from Qwen's latest model — Qwen3.8-27B. We generated this dataset locally by running the model on our own infrastructure. It covers 4 domains with prompts sourced from 12 diverse open-source datasets. Dataset Overview Metric Value Total Examples 40,000 Teacher Model Qwen3.8-27B Model Precision FP8 Reasoning Effort medium… See the full description on the dataset page: https://huggingface.co/datasets/ArkhAngelLifeJiggy/Qwen3.8-27B-Distillation-40K.

sourceHugging Faceapache-2.0updated 29d agoView on Hugging Face

1likes157downloads

Dataset Card

[image]

Qwen3.8-27B-Distillation (40K Traces)

Qwen3.8-27B-Distillation is a dataset containing 40,000 reasoning traces distilled from Qwen's latest model — Qwen3.8-27B. We generated this dataset locally by running the model on our own infrastructure. It covers 4 domains with prompts sourced from 12 diverse open-source datasets.

Dataset Overview

Metric	Value
Total Examples	40,000
Teacher Model	Qwen3.8-27B
Model Precision	FP8
Reasoning Effort	medium
Total Tokens	229,980,687 (~230M)
Assistant Tokens	224,414,445 (~224.4M)
Thinking Tokens	147,744,072 (~147.7M)
User Tokens	5,566,242 (~5.57M)
Avg Tokens per Example	~5,749
Inference Engine	vLLM
Estimated Generation Cost	~$80

Note: The dataset underwent basic deduplication: removal of duplicate and invalid examples.

Domains

Domain	Examples	Percentage
code	16,006	40.02%
math	10,994	27.48%
science	7,000	17.50%
logic	6,000	15.00%

Sources

Prompts were sampled from 12 open-source datasets:

Source	Examples	Percentage
ianncity/Hunter-Alpha-Programming-160000x	10,014	25.04%
ianncity/GLM-5.2-Logic-Puzzles	5,500	13.75%
ianncity/GLM-5.2-Science	5,051	12.63%
open-r1/OpenR1-Math-220k	4,000	10.00%
nvidia/OpenMathReasoning	3,994	9.98%
Modotte/CodeX-2M-Thinking	2,999	7.50%
GetSoloTech/Code-Reasoning	2,993	7.48%
AI-MO/NuminaMath-CoT	2,000	5.00%
EricLu/SCP-116K	1,400	3.50%
zwhe99/DeepMath-103K	1,000	2.50%
Hothan/OlympiadBench	549	1.37%
lukaemon/bbh	500	1.25%
Total	40,000	100%

Structure

Each example is a single JSONL line with the following schema:

json

{
  "messages": [
    {"role": "user", "content": "<prompt>"},
    {"role": "assistant", "content": "<think>...</think>\n\n<response>"}
  ],
  "domain": "code | math | science | logic",
  "category": "python | physics | logic_puzzle | ...",
  "source": "ianncity/Hunter-Alpha-Programming-160000x | ...",
  "tokens_total": 5749,
  "tokens_think": 3200
}

Field	Type	Description
`messages`	list[dict]	Standard chat format with `user` and `assistant` roles
`messages[1].content`	str	Always wrapped in `<think>...</think>` tags followed by the final response
`domain`	str	One of: `code`, `math`, `science`, `logic`
`category`	str	Fine-grained topic (e.g. `python`, `physics`, `olympiads`) or `-` if unspecified
`source`	str	Original dataset the prompt was sampled from
`tokens_total`	int	Total token count (user + assistant), counted with Qwen3.8 tokenizer
`tokens_think`	int	Token count inside `<think>` block only

Usage

With Hugging Face Datasets

python

from datasets import load_dataset

dataset = load_dataset("faunix/Qwen3.8-27B-Distillation-40K", split="train")

print(dataset[0]["messages"])   # full chat with <think> + response
print(dataset[0]["domain"])     # code / math / science / logic
print(dataset[0]["tokens_think"])  # thinking tokens count

With Axolotl

yaml

datasets:
  - path: faunix/Qwen3.8-27B-Distillation-40K
    type: chat_template
    chat_template: qwen3_5
    split: train

With Unsloth

python

from unsloth import FastLanguageModel
from datasets import load_dataset

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Qwen3.5-9B",
    max_seq_length=8192,
)

dataset = load_dataset("faunix/Qwen3.8-27B-Distillation-40K", split="train")

Sponsor

This dataset would not exist without the help of lium.io! By granting us GPU access, we were able to create this dataset, and we will create more, and train many more interesting models! Therefore, as a sign of gratitude, we mention them here! :)

We are on...

![Follow on X](https://x.com/Faunix_AI)

And also subscribe to our HF organization! This way you won't miss new models and other projects we will be publishing! :)

☕ Support Faunix :)

We are two AI developers: **Didiblud** (12 years old) and **Limen4ik** (13 years old). Your support is very important to us — just like everyone else, we want to eat, so we are leaving our donation links below!

—💳 Card / Apple Pay / PayPal: ⭐ Recommended! — app.lava.top/faunix_ai
—💎 TON: UQArgd4zYOA0I5QSK7ylSlvgyaE5Md8pm-O-RyvSnTTk8Me0
—💵 USDT (TRC-20): TVCgvfbrUQC1nKD6cHcSP7JGirEEx5f4Te

(Please note: 60% goes into our personal pockets fund, and 40% goes directly to Faunix development — buying domains, GPU hours, etc.)

Note

This dataset is open-source and distributed under the Apache 2.0 license, so you can use it in experiments and fine-tuning of your own models! We look forward to seeing what models you can fine-tune based on our dataset! :)