tunachiu's picture
Upload model
11e6fb0 verified
|
Raw History Blame Contribute Delete
960 Bytes
metadata
base_model: Qwen/Qwen3.5-4B
library_name: peft
language: vi
tags:
  - lora
  - asr-correction

llm_adapter_text_correction

LoRA adapter (rank 8, alpha 16) for Qwen/Qwen3.5-4B that corrects the ASR transcript of one user turn.

  • Input (user message): conversation context (time, location, profile), earlier turns, the ASR text, then an instruction to fix spelling errors only, staying close to the ASR text; it may add "." and "," between clauses, capitalise proper names and write digits or symbols (times, amounts, phone numbers, emails).
  • Output: the corrected text only. Use with thinking off (chat_template_kwargs={"enable_thinking": false}), greedy decoding.
  • Training: 20k pairs (ASR text -> true text), real transcripts plus synthetic ASR noise; 1 epoch, lr 2e-4.

Serving with vLLM as a LoRA module:

vllm serve Qwen/Qwen3.5-4B --enable-lora --max-lora-rank 64 --lora-modules text_correction=tunachiu/llm_adapter_text_correction