RichardErkhov/Lyte_-_Llama-3.2-3B-Overthinker-4bits

Quantization made by Richard Erkhov.

Github

Discord

Request more models

Llama-3.2-3B-Overthinker - bnb 4bits

• Model creator: https://huggingface.co/Lyte/
• Original model: https://huggingface.co/Lyte/Llama-3.2-3B-Overthinker/

Original model description:

language: - en license: apache-2.0 tags: - text-generation-inference - transformers - unsloth - llama - trl - sft base_model: unsloth/llama-3.2-3b-instruct-bnb-4bit datasets: - Lyte/Reasoning-Paused pipeline_tag: text-generation model-index: - name: Llama-3.2-3B-Overthinker results: - task: type: text-generation name: Text Generation dataset: name: IFEval (0-Shot) type: HuggingFaceH4/ifeval args: num_few_shot: 0 metrics: - type: inst_level_strict_acc and prompt_level_strict_acc value: 64.08 name: strict accuracy source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Lyte/Llama-3.2-3B-Overthinker name: Open LLM Leaderboard - task: type: text-generation name: Text Generation dataset: name: BBH (3-Shot) type: BBH args: num_few_shot: 3 metrics: - type: acc_norm value: 20.1 name: normalized accuracy source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Lyte/Llama-3.2-3B-Overthinker name: Open LLM Leaderboard - task: type: text-generation name: Text Generation dataset: name: MATH Lvl 5 (4-Shot) type: hendrycks/competition_math args: num_few_shot: 4 metrics: - type: exact_match value: 2.64 name: exact match source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Lyte/Llama-3.2-3B-Overthinker name: Open LLM Leaderboard - task: type: text-generation name: Text Generation dataset: name: GPQA (0-shot) type: Idavidrein/gpqa args: num_few_shot: 0 metrics: - type: acc_norm value: 1.23 name: acc_norm source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Lyte/Llama-3.2-3B-Overthinker name: Open LLM Leaderboard - task: type: text-generation name: Text Generation dataset: name: MuSR (0-shot) type: TAUR-Lab/MuSR args: num_few_shot: 0 metrics: - type: acc_norm value: 3.9 name: acc_norm source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Lyte/Llama-3.2-3B-Overthinker name: Open LLM Leaderboard - task: type: text-generation name: Text Generation dataset: name: MMLU-PRO (5-shot) type: TIGER-Lab/MMLU-Pro config: main split: test args: num_few_shot: 5 metrics: - type: acc value: 22.06 name: accuracy source: url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=Lyte/Llama-3.2-3B-Overthinker name: Open LLM Leaderboard

Model Overview:

• Training Data: This model was trained on a dataset with columns for initial reasoning, step-by-step thinking, verifications after each step, and final answers based on full context. Is it better than the original base model? Hard to say without proper evaluations, and I don’t have the resources to run them manually.

• Context Handling: The model benefits from larger contexts (minimum 4k up to 16k tokens, though it was trained on 32k tokens). It tends to "overthink," so providing a longer context helps it perform better.

• Performance: Based on my very few manual tests, the model seems to excel in conversational settings—especially for mental health, creative tasks and explaining stuff. However, I encourage you to try it out yourself using this Colab Notebook.

• Dataset Note: The publicly available dataset is only a partial version. The full dataset was originally designed for a custom Mixture of Experts (MoE) architecture, but I couldn't afford to run the full experiment.

• Acknowledgment: Special thanks to KingNish for reigniting my passion to revisit this project. I almost abandoned it after my first attempt a month ago. Enjoy this experimental model!

Inference Code:

• Feel free to make the steps and verifications collapsable and the initial reasoning too, you can show only the final answer to get an o1 feel(i don't know)
• Note: A feature we have here is the ability to control how many steps and verifications you want.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Lyte/Llama-3.2-3B-Overthinker"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")

def generate_response(prompt, max_tokens=16384, temperature=0.8, top_p=0.95, repeat_penalty=1.1, num_steps=3):
    messages = [{"role": "user", "content": prompt}]

    # Generate reasoning
    reasoning_template = tokenizer.apply_chat_template(messages, tokenize=False, add_reasoning_prompt=True)
    reasoning_inputs = tokenizer(reasoning_template, return_tensors="pt").to(model.device)

    reasoning_ids = model.generate(
        **reasoning_inputs,
        max_new_tokens=max_tokens // 3,
        temperature=temperature,
        top_p=top_p,
        repetition_penalty=repeat_penalty
    )
    reasoning_output = tokenizer.decode(reasoning_ids[0, reasoning_inputs.input_ids.shape[1]:], skip_special_tokens=True)
    
    # Generate thinking (step-by-step and verifications)
    messages.append({"role": "reasoning", "content": reasoning_output})
    thinking_template = tokenizer.apply_chat_template(messages, tokenize=False, add_thinking_prompt=True, num_steps=num_steps)
    thinking_inputs = tokenizer(thinking_template, return_tensors="pt").to(model.device)

    thinking_ids = model.generate(
        **thinking_inputs,
        max_new_tokens=max_tokens // 3,
        temperature=temperature,
        top_p=top_p,
        repetition_penalty=repeat_penalty
    )
    thinking_output = tokenizer.decode(thinking_ids[0, thinking_inputs.input_ids.shape[1]:], skip_special_tokens=True)
    
    # Generate final answer
    messages.append({"role": "thinking", "content": thinking_output})
    answer_template = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    answer_inputs = tokenizer(answer_template, return_tensors="pt").to(model.device)

    answer_ids = model.generate(
        **answer_inputs,
        max_new_tokens=max_tokens // 3,
        temperature=temperature,
        top_p=top_p,
        repetition_penalty=repeat_penalty
    )
    answer_output = tokenizer.decode(answer_ids[0, answer_inputs.input_ids.shape[1]:], skip_special_tokens=True)
    return reasoning_output, thinking_output, answer_output

# Example usage:
prompt = "Explain the process of photosynthesis."
response = generate_response(prompt, num_steps=5)

print("Response:", response)

Uploaded model

• Developed by: Lyte
• License: apache-2.0
• Finetuned from model : unsloth/llama-3.2-3b-instruct-bnb-4bit

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

Notice:

• The problem with runnning evals is that they won't make use of the correct template and it won't be a true eval then would it? so these barely test the model.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	19.00
IFEval (0-Shot)	64.08
BBH (3-Shot)	20.10
MATH Lvl 5 (4-Shot)	2.64
GPQA (0-shot)	1.23
MuSR (0-shot)	3.90
MMLU-PRO (5-shot)	22.06