Instructions to use sajjadiba/urdu-asr-error-correction-tiny-aya-fire with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use sajjadiba/urdu-asr-error-correction-tiny-aya-fire with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-fire")
model = PeftModel.from_pretrained(base_model, "sajjadiba/urdu-asr-error-correction-tiny-aya-fire")

Transformers

How to use sajjadiba/urdu-asr-error-correction-tiny-aya-fire with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="sajjadiba/urdu-asr-error-correction-tiny-aya-fire")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("sajjadiba/urdu-asr-error-correction-tiny-aya-fire", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use sajjadiba/urdu-asr-error-correction-tiny-aya-fire with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "sajjadiba/urdu-asr-error-correction-tiny-aya-fire"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sajjadiba/urdu-asr-error-correction-tiny-aya-fire",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/sajjadiba/urdu-asr-error-correction-tiny-aya-fire

SGLang

How to use sajjadiba/urdu-asr-error-correction-tiny-aya-fire with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "sajjadiba/urdu-asr-error-correction-tiny-aya-fire" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sajjadiba/urdu-asr-error-correction-tiny-aya-fire",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "sajjadiba/urdu-asr-error-correction-tiny-aya-fire" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sajjadiba/urdu-asr-error-correction-tiny-aya-fire",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use sajjadiba/urdu-asr-error-correction-tiny-aya-fire with Docker Model Runner:
```
docker model run hf.co/sajjadiba/urdu-asr-error-correction-tiny-aya-fire
```

Urdu ASR Error Correction (Tiny-Aya-Fire LoRA Adapter)

This repository contains fine-tuned LoRA adapter weights for post-ASR error correction and sentence segmentation of Urdu transcripts generated by openai/whisper-large-v3-turbo.

Base Model: CohereLabs/tiny-aya-fire
Dataset: urdu-asr-error-correction-data
License: CC BY-NC 4.0

Quickstart / Usage

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

base_model_id = "CohereLabs/tiny-aya-fire"
adapter_id = "your-username/urdu-asr-error-correction-tiny-aya-fire"

# Load Base Model & Tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id, 
    torch_dtype=torch.bfloat16, 
    device_map="auto"
)

# Load Fine-Tuned Adapter
model = PeftModel.from_pretrained(base_model, adapter_id)

# Zero-Shot Prompt format as defined in the paper
ZERO_SHOT_PROMPT = """Task: Correct Urdu ASR errors.
Rules:
1. Fix ONLY spelling and character errors.
2. Keep correct words EXACTLY as is --- DO NOT rewrite or paraphrase.
3. DO NOT add any notes, explanations, or markers.
4. STOP immediately after the corrected sentence.

Input: {input_text}
Corrected:"""

# Example Raw ASR Output from Whisper
raw_asr_input = "آپ کا ان پٹ اردو جملہ یہاں درج کریں"
prompt = ZERO_SHOT_PROMPT.format(input_text=raw_asr_input)

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=128)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation

If you use this model adapter in your research, please cite our paper:

@inproceedings{haider2026improving,
  title={Improving Urdu ASR with Fine-Tuned Small Language Models for Post-ASR Error Correction},
  author={Haider, Sajjad and Shaikh, Solat Jabeen and Aqib, Zuha and Inayat, Farah and Ahmed, Zehra},
  booktitle={Proceedings of the 2nd Workshop on Speech and Audio Language Models (SALMA @ EMNLP 2026)},
  year={2026}
  organization={Artificial Intelligence Lab, Department of Computer Science, Institute of Business Administration (IBA), Karachi, Pakistan}
}

Downloads last month: 36

Model tree for sajjadiba/urdu-asr-error-correction-tiny-aya-fire

Base model

CohereLabs/tiny-aya-base

Finetuned

CohereLabs/tiny-aya-fire

Adapter

(5)

this model