mongolian-qwen-7b-chatbot

mongolian-qwen-7b-chatbot

Description

Qwen 2.5 7B fine-tuned on Mongolian news data for chatbot

Model Details

• Base Model: Qwen/Qwen2.5-7B-Instruct
• Language: Mongolian (mn)
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Training Data: Eduge Mongolian News Dataset (75,000+ articles)

Training Configuration

• LoRA Rank: 32
• LoRA Alpha: 64
• Epochs: 3
• Learning Rate: 2e-4
• Batch Size: 4
• Max Sequence Length: 1024

Mongolian Tokens Added

• Total new tokens: ~9,500
• Sources: Mongolian-NLP repository
  • Most frequent words
  • Abbreviations
  • District/place names
  • Country names
  • Named entities (NER)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("ErkaMarka/mongolian-qwen-7b-chatbot")

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Resize embeddings for new tokens
base_model.resize_token_embeddings(len(tokenizer))

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "ErkaMarka/mongolian-qwen-7b-chatbot")

# Generate
messages = [{"role": "user", "content": "Монгол улсын нийслэл хот юу вэ?"}]
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=150)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Evaluation Results

Evaluated on 100 Mongolian Q&A pairs using BLEU score.

License

Apache 2.0

Citation

@misc{mongolian_qwen_7b_chatbot},
  author = {Your Name},
  title = {mongolian-qwen-7b-chatbot},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/ErkaMarka/mongolian-qwen-7b-chatbot}
}