CreitinGameplays/Llama-3.2-3B-Instruct-R1-v1

Llama 3.1 8B R1 Experimental

Chat template format:

<|start_header_id|>system<|end_header_id|>
You are focused on providing systematic, well-reasoned responses. Response Structure: - Format: <think>{{reasoning}}</think>{{answer}} - Reasoning: Minimum 6 logical steps only when it required in <think> block - Process: Think first, then answer.

You are a helpful AI assistant named Llama, made by Meta AI.
<|eot_id|><|start_header_id|>user<|end_header_id|>

How many r's are in strawberry?<|eot_id|><|start_header_id|>assistant<|end_header_id|><think>

Run this model:

# test the model
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, TextStreamer

def main():
    model_id = "CreitinGameplays/Llama-3.2-3B-Instruct-R1-v1"

    # Load the tokenizer.
    tokenizer = AutoTokenizer.from_pretrained(model_id, add_eos_token=True)

    # Load the model using bitsandbytes 8-bit quantization if CUDA is available.
    if torch.cuda.is_available():
        model = AutoModelForCausalLM.from_pretrained(
            model_id,
            load_in_8bit=True,
            device_map="auto"
        )
        device = torch.device("cuda")
    else:
        model = AutoModelForCausalLM.from_pretrained(model_id)
        device = torch.device("cpu")

    # Define the generation parameters.
    generation_kwargs = {
        "max_new_tokens": 2048,
        "do_sample": True,
        "temperature": 0.5,
        "top_p": 0.9,
        "repetition_penalty": 1.1,
        "num_return_sequences": 1,
        "forced_eos_token_id": tokenizer.eos_token_id,
        "pad_token_id": tokenizer.eos_token_id
    }

    print("Enter your prompt (type 'exit' to quit):")
    while True:
        # Get user input.
        user_input = input("Input> ")
        if user_input.lower().strip() in ("exit", "quit"):
            break

        # Construct the prompt in your desired format.
        prompt = f"""
<|start_header_id|>system<|end_header_id|>
You are focused on providing systematic, well-reasoned responses. Response Structure: - Format: <think>{{reasoning}}</think>{{answer}} - Reasoning: Minimum 6 logical steps only when it required in <think> block - Process: Think first, then answer.

You are a helpful AI assistant named Llama, made by Meta AI.
<|eot_id|><|start_header_id|>user<|end_header_id|>

How many r's are in strawberry?<|eot_id|><|start_header_id|>assistant<|end_header_id|><think>
"""

        # Tokenize the prompt and send to the selected device.
        input_ids = tokenizer.encode(prompt, return_tensors="pt", add_special_tokens=True).to(device)

        # Create a new TextStreamer instance for streaming responses.
        streamer = TextStreamer(tokenizer)
        generation_kwargs["streamer"] = streamer

        print("\nAssistant Response:")
        # Generate the text (tokens will stream to stdout via the streamer).
        outputs = model.generate(input_ids, **generation_kwargs)

if __name__ == "__main__":
    main()

Or alternatively:

import torch
from transformers import pipeline

model_id = "CreitinGameplays/Llama-3.2-3B-Instruct-R1-v1"

pipe = pipeline(
    "text-generation",
    model=model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [{"role": "user", "content": "hello there!"}]

outputs = pipe(
    messages,
    temperature=0.5,
    repetition_penalty=1.1,
    max_new_tokens=2048
)

print(outputs[0]["generated_text"][-1])