Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,106 @@
+---
+language:
+- zh
+license: apache-2.0
+base_model: unsloth/Qwen3.5-9B
+tags:
+- lora
+- peft
+- dpo
+- text-generation
+- chinese
+- humanization
+- academic-writing
+pipeline_tag: text-generation
+---
+
+# Qwen3.5-9B Humanize DPO v20
+
+LoRA adapter fine-tuned with **DPO** on Qwen3.5-9B for Chinese text humanization. This is the **latest and most capable version**, especially strong on academic and technical Chinese text. It uses v18-75 model outputs as rejected samples, teaching the model to correct its own remaining weaknesses.
+
+## Model Details
+
+| Item | Value |
+|---|---|
+| Base model | `unsloth/Qwen3.5-9B` |
+| Starting point | DPO v18, checkpoint-75 |
+| Fine-tuning method | DPO |
+| LoRA rank | 16 |
+| Training data | 2000 pairs (chosen = CSL human text, rejected = v18-75 model outputs) |
+| Training steps | 250 steps (2 epochs) |
+| Final loss | 0.34 |
+| Final margin | ~1.5–2.5 |
+| Final accuracy | ~93–100% |
+
+## What It Does
+
+The most natural-sounding version in the series, particularly on academic and technical texts:
+
+- **Academic papers**: rewrites while preserving all numbers, model names, formulas, and technical terms — verified on 10 academic scenarios with 3 samples each
+- **Technical reports**: maintains technical register without sounding robotic
+- **Daily text**: more concise and natural than previous versions
+
+## Usage
+
+```python
+from unsloth import FastLanguageModel
+from peft import PeftModel
+
+base_model, proc = FastLanguageModel.from_pretrained(
+    "unsloth/Qwen3.5-9B", max_seq_length=2048, load_in_4bit=False,
+)
+tokenizer = proc.tokenizer if hasattr(proc, "tokenizer") else proc
+
+model = PeftModel.from_pretrained(
+    base_model, "XiangJinYu/Qwen3.5-9B-Humanize-DPO-v20", is_trainable=False,
+)
+if hasattr(model, "config") and getattr(model.config, "model_type", "") == "qwen3_5":
+    model.config.model_type = "qwen3"
+FastLanguageModel.for_inference(model)
+
+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。"
+text = "本文提出了一种基于U-Net改进的医学影像分割方法，Dice系数达到0.923，较基线方法提升了4.7个百分点，推理速度提升约30%。"
+messages = [{"role": "user", "content": [{"type": "text", "text": f"{instruction}\n\n原文：{text}"}]}]
+prompt = tokenizer.apply_chat_template(
+    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
+)
+inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+# Recommended: temperature=0.6-0.65 for academic texts to avoid rare term substitution errors
+outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.65,
+                         top_p=0.9, do_sample=True, repetition_penalty=1.1)
+gen = outputs[0][inputs["input_ids"].shape[1]:]
+print(tokenizer.decode(gen, skip_special_tokens=True))
+```
+
+## Training Details
+
+- **DPO data**: 2000 pairs — chosen = CSL human academic text, rejected = v18-75 model outputs (pure self-play, no casual mixing)
+- **Reference model**: v18 checkpoint-75 (same as starting point)
+- **β**: 0.2, lr=5e-7 (half of v19), 2 epochs (250 steps)
+- **Key design choices**:
+  - Lower LR (5e-7 vs 1e-6 in v19) for more stable convergence
+  - Pure self-play rejected (no casual text mixing) eliminates gradient variance from mixed data types
+  - Rejected reward stayed negative and decreasing throughout all 250 steps — no training instability
+
+## Academic Test Results
+
+Tested on 10 academic scenarios (3 samples each), all key numbers preserved:
+
+| Scenario | Numbers verified |
+|---|---|
+| NLP paper abstract | BLEU +3.2%, complexity -15% ✅ |
+| Medical image segmentation | Dice 0.923, +4.7%, speed +30% ✅ |
+| Graph neural network | O(n log n), F1 +2.8%, time -40% ✅ |
+| SPWM inverter | 83.9%, 81.9%, 6~18V, IEC 61000-4-2 ✅ |
+| Embedded system test | ±0.5LSB, 8ms, 28mW, 0.6mW ✅ |
+
+> **Note**: Use temperature 0.60–0.65 for academic texts. Higher temperatures occasionally cause rare technical term substitutions.
+
+## Model Series
+
+| Model | Type | Recommended for |
+|---|---|---|
+| [SFT v15](https://huggingface.co/XiangJinYu/Qwen3.5-9B-Humanize-SFT-v15) | SFT | Foundation |
+| [v16](https://huggingface.co/XiangJinYu/Qwen3.5-9B-Humanize-DPO-v16) | DPO | General use, balanced |
+| [v18](https://huggingface.co/XiangJinYu/Qwen3.5-9B-Humanize-DPO-v18) | DPO | Stable clean baseline |
+| **This model** | DPO | ✅ Academic/technical, latest |