Overview

The bingbangboom/Qwen3006B-transcriber-beta is a specialized Qwen3-0.6B model developed by bingbangboom, primarily designed as a post-processor for local Automatic Speech Recognition (ASR) systems. It was fine-tuned from unsloth/qwen3-0.6b-unsloth-bnb-4bit and leverages Unsloth and Huggingface's TRL library for efficient training.

Key Capabilities

• ASR Post-Processing: Its core function is to refine and process raw transcripts generated by local ASR systems.
• Optimized Training: Benefits from Unsloth's accelerated training methods, indicating potential for efficient deployment.
• Flexible Deployment: Available in various GGUF quantization formats (Qwen3.5-0.8B.F16.gguf, Qwen3.5-0.8B.Q8_0.gguf, Qwen3.5-0.8B.Q5_K_M.ggu, Qwen3.5-0.8B.Q4_K_M.gguf) and as a Lora merged safetensor, catering to different hardware and performance needs.

Recommended Usage

• Prompt Format: Expects the input transcript directly, e.g., {input transcript} for chat or ${output} for Handy integration.
• Inference Settings: Recommended settings include a low temperature (0.1), top_k of 10, top_p of 0.95, min_p of 0.05, and a repeat_penalty of 1.0.
• No System Prompt: The model does not require a system prompt. Users should disable model thinking by adding {%- set enable_thinking = false %} to the Jinja Prompt Template, particularly noted for LMStudio users.