deftwriting_distil_Qwen3.5_9B

DeftWriting Distil — Qwen3.5-9B LoRA

This is a rank-16 LoRA adapter for Qwen/Qwen3.5-9B. It was trained to rewrite text using outputs distilled from DeftWriting.

The training conversation has this form:

• System: Rewrite this text.\n/no_think
• User: source text
• Assistant: DeftWriting rewrite

Training used 6,825 rows for two epochs with a learning rate of 1e-4. A deterministic, category-stratified 10% split (760 rows) was held out from training. The split seed was 42. Full configuration details are in training_manifest.json.

Use

import torch
from peft import PeftModel
from transformers import AutoTokenizer, Qwen3_5ForCausalLM

adapter = "trentmkelly/deftwriting_distil_Qwen3.5_9B"
tokenizer = AutoTokenizer.from_pretrained(adapter)
base = Qwen3_5ForCausalLM.from_pretrained(
    "Qwen/Qwen3.5-9B",
    dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter)

messages = [
    {"role": "system", "content": "Rewrite this text.\n/no_think"},
    {"role": "user", "content": "Text to rewrite."},
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, do_sample=False, max_new_tokens=1024)
print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))

Use enable_thinking=False so inference matches the training template.

Training run

The run is tracked on Weights & Biases.

This adapter has not received a broad safety evaluation. Review generated text before use.