AashishKumar/Cn_2_9_Hinglish_llama3_7b_4kAk

Uploaded model

• Developed by: AashishKumar
• License: apache-2.0
• Finetuned from model : cognitivecomputations/dolphin-2.9-llama3-8b

from unsloth.chat_templates import get_chat_template

# Assuming you've initialized your tokenizer and model
tokenizer = get_chat_template(
    tokenizer,
    chat_template="chatml",  # Adjust as per your template needs
    mapping={"role": "from", "content": "value", "user": "human", "assistant": "gpt"},
    map_eos_token=True,
)

FastLanguageModel.for_inference(model)  # Ensure model is optimized for inference

messages = [
    {"from": "system", "value": "you are assistant designed to talk to answer any user question like a normal human would. Make sure any names are in english"},
    {"from": "human", "value": "mujhe kuch acchi movies recommend kro"}  # Example Hinglish input
]

inputs = tokenizer.apply_chat_template(
    messages,
    truncation=True,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to("cuda")

outputs = model.generate(
        input_ids=inputs, 
        max_new_tokens=64,
        do_sample=True,
        temperature=0.7,
        top_k=50,
        top_p=0.95,
        use_cache=True,
        no_repeat_ngram_size=3,
        num_return_sequences=1
    )
decoded_outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]

print(decoded_outputs)  # Adjust how you handle outputs based on your application needs