flowscribe-qwen2.5-0.5b-v2/README.md

language, license, base_model, tags, pipeline_tag, datasets

language	license	base_model	tags	pipeline_tag	datasets
en	mit	Qwen/Qwen2.5-0.5B-Instruct	text-generation fine-tuned lora gguf speech-to-text text-cleanup unsloth qwen2 conversational	text-generation	Abdullahu5mani/flowscribe-dataset

FlowScribe — Qwen2.5-0.5B Speech Transcript Formatter (v2)

A fine-tuned version of Qwen2.5-0.5B-Instruct that converts raw, messy speech-to-text output into clean, formatted text across multiple writing styles.

GitHub: github.com/Abdullahu5mani/flowscribe

---

The Problem

Voice dictation tools like Whisper produce transcripts full of filler words (um, uh, like), self-corrections (make it 5... no wait, 6), and no punctuation or formatting. This model post-processes those transcripts into polished text, with awareness of the desired output style.

---

Styles

Style	Behavior
Auto	Intelligent default — removes fillers, fixes grammar, handles self-corrections, applies structure
Professional	Formal business tone, structured layout, perfect grammar
Casual	Keeps the speaker's voice, light cleanup, contractions preserved
Verbatim	Preserves exact wording, only strips um/uh and applies spoken formatting commands
Software_Dev	Formats code terms, variable names (camelCase, snake_case), technical jargon
Enthusiastic	High energy, exclamation marks, positive phrasing

---

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Abdullahu5mani/flowscribe-qwen2.5-0.5b-v2"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

def format_transcript(raw_text, style="Auto"):
    messages = [
        {
            "role": "system",
            "content": "You are Flowscribe, an expert Speech-to-Text post-processing AI. You accurately transcribe and format text based on a specific style instruction."
        },
        {
            "role": "user",
            "content": f"Transcribe and format this with style: {style}\nInput: {raw_text}"
        }
    ]
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer([text], return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
    output_ids = outputs[0][len(inputs.input_ids[0]):]
    return tokenizer.decode(output_ids, skip_special_tokens=True)

# Examples
print(format_transcript(
    "um so the meeting is at 5... no wait make it 6 and uh we need to discuss the q3 budget",
    style="Professional"
))
# → "The meeting is at 6 PM to discuss the Q3 budget."

print(format_transcript(
    "the api endpoint is slash api slash users new line it takes a POST request with JSON",
    style="Software_Dev"
))
# → "The API endpoint is `/api/users`\nIt takes a POST request with JSON."

---

GGUF (Quantized) Usage

A Q4_K_M quantized GGUF version is included in this repository for fast CPU/GPU inference via llama-cpp-python.

from llama_cpp import Llama

llm = Llama(
    model_path="model_q4_k_m.gguf",
    n_ctx=2048,
    n_gpu_layers=-1,   # Set to 0 for CPU-only
    verbose=False
)

response = llm.create_chat_completion(
    messages=[
        {
            "role": "system",
            "content": "You are Flowscribe, an expert Speech-to-Text post-processing AI. You accurately transcribe and format text based on a specific style instruction."
        },
        {
            "role": "user",
            "content": "Transcribe and format this with style: Casual\nInput: hey um so i was thinking we could like grab lunch tomorrow you know around noon ish"
        }
    ],
    max_tokens=256,
    temperature=0.1,
)
print(response["choices"][0]["message"]["content"])
# → "Hey, I was thinking we could grab lunch tomorrow around noon."

---

Model Details

Property	Value
Version	v2
Base model	Qwen/Qwen2.5-0.5B-Instruct
Fine-tuning method	LoRA (via Unsloth)
Parameters	~500M (72.4% trained)
Training epochs	3
Learning rate	2e-5
Effective batch size	16 (batch 8 × grad accumulation 2)
Sequence length	2048
Optimizer	AdamW 8-bit
Final training loss	0.616
Training hardware	NVIDIA RTX 4070 Laptop GPU 8GB
Chat template	ChatML
Quantization	Q4_K_M (via llama.cpp)

---

Training Data

Trained on ~27,400 synthetically generated examples from flowscribe-dataset.

Each example is an Alpaca-style JSON object:

{
  "instruction": "Transcribe and format this with style: Professional",
  "input": "um so like the uh proposal is due friday and we need to finalize the, i mean confirm the budget",
  "output": "The proposal is due Friday and we need to confirm the budget."
}

Data was generated using Google Gemini (primary) and 16 free OpenRouter models (fallback) across 10 domain scenarios: business email, software dev, personal messages, productivity lists, medical notes, and more.

---

Limitations

• Optimized for English only
• Training data is synthetic — real-world dictation edge cases may vary
• The 0.5B parameter size prioritizes speed and local deployment over raw capability

---

Files

File	Description
model.safetensors	Full-precision fine-tuned weights (BF16)
model_q4_k_m.gguf	Q4_K_M quantized GGUF for llama.cpp
config.json	Model configuration
tokenizer.json	Tokenizer
chat_template.jinja	ChatML chat template

---

License

MIT — see LICENSE