Subh775/hinglish-finetuned-demo

Hinglish Fine-tuned Conversational Model

This model is fine-tuned on Hinglish conversation data using LoRA adapters. It's designed to respond to queries in Hinglish (a mix of Hindi and English).

Model Details

• Base model: facebook/opt-350m
• Fine-tuning: LoRA adapters
• Training dataset: one-thing/chatbot_arena_conversations_hinglish
• Language: Hinglish (Hindi-English code-mixed)

Usage

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

# Load configuration
config = PeftConfig.from_pretrained("Subh775/hinglish-finetuned-demo")

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Load LoRA model
model = PeftModel.from_pretrained(base_model, "Subh775/hinglish-finetuned-demo")

# Prepare input
prompt = "Human: Explain what is an Artificial Neural Network?\nAssistant:"
inputs = tokenizer(prompt, return_tensors="pt")

# Try with modified generation parameters
outputs = model.generate(
    **inputs,
    max_length=100,
    min_new_tokens=10,  # Force generating at least some new tokens
    temperature=0.86,    # Add some randomness
    top_p=0.9,
    no_repeat_ngram_size=3,  # Avoid repeating trigrams
    repetition_penalty=1.5,  # Penalize repetition more heavily
    do_sample=True     # Use sampling instead of greedy decoding
)

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

Limitations

• The model is fine-tuned on a specific dataset and may not generalize to all Hinglish dialects or topics.
• It works best for conversational queries similar to those in the training data. """