MaxDevv/Qwen3.8-27B-Distill-1M-3.12B-Tokens

Qwen3.8-27B-Distill-1M-4.83B-Tokens

A unified, globally deduplicated, large-scale supervised distillation corpus built from 992,318 conversations generated by Qwen/Qwen3.8-27B, containing 4,834,771,862 target output tokens (3,570,459,498 reasoning tokens + 1,264,312,364 final response tokens) and 5,104,980,053 total sequence tokens.

---

1. Dataset Overview

This dataset merges, aligns, and deduplicates the two primary high-quality Qwen3.8-27B generation corpora on Hugging Face:

• RadixArk/Qwen3.8-27B-Regen (7 splits covering code, chat, math, and STEM reasoning).
• DaoCloud/Qwen3.8-27B-Drafter-SFT (64 Parquet partitions covering multi-source instruction and speculative draft completions).

All records are normalized into a single multi-turn schema with explicit separation between internal chain-of-thought reasoning (reasoning_content) and clean assistant responses (content).

---

2. Token Statistics & Distribution

The statistics below are computed across representative samples from all 10 unified dataset shards using the official Qwen/Qwen3.8-27B BPE tokenizer:

Metric Category	Mean	Median	Mode	P80 (80th %)	P95 (95th %)	Max	Total Volume
Input / Prompt Tokens	272.3	115.0	10	263.2	795.3	19,584	270,208,191 Tokens
Reasoning Tokens (<think>)	3,598.1	512.5	32,769	4,126.2	22,665.2	58,245	3,570,459,498 Tokens
Final Completion Tokens	1,274.1	230.5	0	922.0	5,990.5	40,118	1,264,312,364 Tokens
Total Output Tokens	4,872.2	1,224.5	32,769	6,746.4	26,502.7	58,696	4,834,771,862 Tokens
Total Sequence Tokens	5,144.4	1,430.5	204	7,186.8	26,968.8	58,935	5,104,980,053 Tokens

Note: The reasoning token mode reflects the 32k context target ceiling common in long-horizon synthetic trajectories.

---

3. Dataset Structure & Schema

The dataset is partitioned into 10 compressed Zstandard JSONL shards (data/qwen3.8-distill-train-*.jsonl.zst). Each record contains:

{
  "id": "daocloud:opb_c19dce9ce4dc923c1461538ceff3beed756587177289cffcddff8e66d90f24a7_gen1",
  "source": "daocloud/opb",
  "conversations": [
    {
      "role": "system",
      "content": "Reasoning effort is set to medium. Balance thoroughness and conciseness in your thinking process."
    },
    {
      "role": "user",
      "content": "Write a python function to compute the longest common subsequence..."
    },
    {
      "role": "assistant",
      "content": "Here is the dynamic programming implementation for Longest Common Subsequence...",
      "reasoning_content": "We need to find the LCS of two strings. Let's analyze the recurrence relation..."
    }
  ],
  "regen_contract": {
    "reasoning_effort": "medium",
    "primary_id": "opb_c19dce9ce4dc923c1461538ceff3beed756587177289cffcddff8e66d90f24a7",
    "continue_final_message": false
  },
  "status": "completed",
  "regeneration": {
    "model": "Qwen/Qwen3.8-27B",
    "source_row_id": "opb_c19dce9ce4dc923c1461538ceff3beed756587177289cffcddff8e66d90f24a7_gen1"
  }
}

---

4. Processing & Deduplication Methodology

1. Token Ingestion & Batch Decoding: Pre-tokenized sequences (input_ids) from DaoCloud Parquet shards were batch-decoded using the official Qwen3.8 tokenizer C++ bindings.
2. ChatML Extraction: Regularized parsing extracted system, user, and assistant turns, isolating internal <think>...</think> traces into reasoning_content and response bodies into content.
3. Cryptographic Deduplication: Cryptographic SHA-256 hashes of normalized user prompts pruned duplicate instructions across upstream overlaps (open-perfectblend, opencodeinstruct, aya, nemotron).
4. Streaming Compression: Formatted samples were streamed into 100,000-record compressed Zstandard archives (.jsonl.zst).

---

5. Provenance & Citations

If you use this unified dataset, please cite the upstream source datasets and models:

@dataset{qwen3.8_distill_1m_4.83b_tokens_2026,
  title={Qwen3.8-27B-Distill-1M-4.83B-Tokens: Unified Distillation and Reasoning Corpus},
  year={2026},
  publisher={Hugging Face}
}

@dataset{daocloud_qwen3.8_drafter_2025,
  title={Qwen3.8-27B Drafter SFT Corpus},
  author={DaoCloud AI Team},
  year={2025},
  url={https://huggingface.co/datasets/DaoCloud/Qwen3.8-27B-Drafter-SFT}
}

@dataset{radixark_qwen3.8_regen_2025,
  title={Qwen3.8-27B Regen Mixture},
  author={RadixArk Team},
  year={2025},
  url={https://huggingface.co/datasets/RadixArk/Qwen3.8-27B-Regen}
}

---

6. Quickstart Usage

from datasets import load_dataset

# Load the complete unified dataset
dataset = load_dataset("json", data_files="data/qwen3.8-distill-train-*.jsonl.zst", split="train")

print(f"Loaded {len(dataset):,} training conversations")
print("Sample turn:", dataset[0]["conversations"])