silas114514/PMTX1-2B-adapter

PMTX1-2B Adapter

PMTX1-2B Adapter is a LoRA adapter trained on top of Qwen/Qwen3.5-2B for Chinese prompt-rewriting coaching (train_aligned style).

Repository Type

This is an adapter repo (PEFT), not a merged full-weight repo.

• Base model required: Qwen/Qwen3.5-2B
• Main weight file: adapter_model.safetensors
• Companion files: adapter_config.json, tokenizer*, chat_template.jinja

One-Command Pull (Transformers)

pip install -U torch transformers peft bitsandbytes accelerate
python -c "from huggingface_hub import snapshot_download; snapshot_download('silas114514/PMTX1-2B-adapter')"

Quick Inference (Transformers + PEFT)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

base_model = "Qwen/Qwen3.5-2B"
adapter_repo = "silas114514/PMTX1-2B-adapter"

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
)

tok = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
if tok.pad_token is None:
    tok.pad_token = tok.eos_token or tok.unk_token

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    quantization_config=bnb,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_repo)
model.eval()

prompt = "你是 Prompt Evolution 的提示词纠偏教练。请只做提示词优化，不要直接代做任务。\n原始提示词：写周报，你看着办就行，快一点。"
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(tok.decode(out[0], skip_special_tokens=True))

vLLM Notes

Use this adapter with a PEFT-aware serving path. If you need pure one-command deploy for broad testers, prefer the merged repo: silas114514/PMTX1-2B-merged.

Ollama Notes

Official qwen3.5 base models can be pulled in Ollama, but this LoRA adapter is not a direct ollama pull/run artifact. Converting custom Qwen3.5 fine-tunes to Ollama-compatible format may require extra conversion support and verification.

Training Snapshot

• Prompt style: train_aligned
• Max steps: 100
• LoRA layers: top 8
• LoRA targets: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
• See run_summary.json, run_config.json, metrics_summary.json

License and Compliance

• This adapter is a derivative work on top of Qwen/Qwen3.5-2B.
• Follow the base model license and usage terms from the upstream repository.

Known Limitations

• Output format can still drift in some prompts; structured post-processing may still be needed in strict production settings.
• Not designed for factual content generation; optimized for prompt rewriting guidance.