XiangJinYu/Qwen3.5-9B-Humanize-Dataset

Chinese Text Humanization Training Data

Training data for the Qwen3.5-9B-Humanize model series — covering the full training pipeline from SFT to DPO alignment.

Task: Rewrite AI-generated Chinese text to sound more like natural human writing, preserving all facts and technical details.

---

Files

File	Samples	Stage	Description
v15_sft_18k.jsonl	18,000	SFT	Supervised fine-tuning data. Chinese academic abstracts from CSL; chosen = human original, rejected = AI-rewritten version. Format: {instruction, input, output, meta}
v16_dpo_4000_formal2000_casual2000.jsonl	4,000	DPO v16	First DPO stage. 2,000 pairs with over-formal rejected + 2,000 pairs with over-casual rejected. Teaches the model to target the human writing distribution rather than shift in one direction. Format: {prompt, input, chosen, rejected, rejected_type, meta}
v17next_dpo_1500_clean_no_ckpt.jsonl	1,500	DPO v18	Second DPO stage (used for v18 training). Clean external rejected only — 1,000 multi-step paraphrase rejected + 500 casual rejected. No model-generated outputs mixed in. Format: {prompt, input, chosen, rejected, rejected_type, meta}
v20_dpo_2000_v18rejected.jsonl	2,000	DPO v20	Third DPO stage. Pure self-play: chosen = CSL human originals, rejected = outputs from v18/checkpoint-75. Trains the model to surpass its own v18-75 baseline. Format: {prompt, input, chosen, rejected, rejected_type, meta}

---

Training Pipeline

Qwen3.5-9B (base)
    └── SFT v15  [v15_sft_18k.jsonl]
          └── DPO v16/ckpt-200  [v16_dpo_4000_formal2000_casual2000.jsonl]
                └── DPO v18/ckpt-75  [v17next_dpo_1500_clean_no_ckpt.jsonl]
                      └── DPO v20/final  [v20_dpo_2000_v18rejected.jsonl]

---

Data Format

SFT (v15_sft_18k.jsonl):

{
  "instruction": "请将下面文本改写得更像自然人写作，保持原意与事实，不要加标题或说明。",
  "input": "<AI-style Chinese text>",
  "output": "<human-style rewrite>",
  "meta": {}
}

DPO (all other files):

{
  "prompt": "请将下面文本改写得更像自然人写作，保持原意与事实，不要加标题或说明。",
  "input": "<AI-style Chinese text>",
  "chosen": "<human original / high-quality rewrite>",
  "rejected": "<AI rewrite / model output / over-formal or over-casual version>",
  "rejected_type": "formal | casual | multistep | ckpt_generated",
  "meta": {}
}

---

Key Design Decisions

• chosen is always the human original: All chosen samples are human-authored Chinese academic abstracts from CSL. The model learns to target the actual human writing distribution, not a curated "better AI" output.
• No casual drift: v16 uses bidirectional rejected (both too-formal and too-casual) so the model learns the center of the human distribution rather than shifting toward colloquial speech.
• Self-play for final stage: v20 uses the v18-75 model's own outputs as rejected. This gives the model a concrete, achievable improvement target and avoids gradient noise from heterogeneous rejected sources.
• Early stopping is critical: v18 training was stopped at step 75 (not final). Rejected reward drifting positive is a key early-stopping signal for DPO.

---

Models Trained on This Data

Model	HuggingFace
SFT v15	XiangJinYu/Qwen3.5-9B-Humanize-SFT
DPO v16	XiangJinYu/Qwen3.5-9B-Humanize-DPO-Round1
DPO v18	XiangJinYu/Qwen3.5-9B-Humanize-DPO-Round2
DPO v20	XiangJinYu/Qwen3.5-9B-Humanize-DPO

---

Source

Text content is derived from the CSL (Chinese Scientific Literature) dataset. CSL contains abstracts from Chinese academic journals and conference papers.