Llama3.1-deep-o1

Llama3.1-deep-o1

• A robust merge of DeepSeek R1 distilled and O1-style long chain-of-thought (CoT) large language models (LLMs).
• It can generate long, coherent solutions and excels at problem-solving tasks among models with 8 billion parameters.

Model Overview

• Supports long reasoning and non-reasoning modes.
• In reasoning mode, generates the thought process for problem solving.
• Suitable for creating solution outlines and analyzing problems
• Or as a foundation model for finetuning and merging.

[!NOTE]
To use the long CoT reasoning mode, use a system prompt like

Explain your reasoning step-by-step using <think>...</think>, then give the final answer inside <response>...</response>.

Examples to Try:

• Write the equations for glycolysis and pyruvate oxidation.
• Calculate net ATP formation from glucose metabolism (excluding electron transport chain).
• Integrate x^2 e^x dx.
• Prove that the complete bipartite graph K_{3,3} isn't planar.
• Derive a formula for the critical angle between two media with refractive indices n_1 and n_2.
• Compare steam vs. diesel engines including their capabilities and historical significance.

Limitations

[!WARNING]
This model is experimental. While the model provides coherent and expert-like responses, users should verify its outputs for accuracy - especially in calculations or logical reasoning tasks.

• It is not optimized for conversational tasks but performs well in single-turn question answering.
• Inconsistent formatting for mathematical equations and LaTeX code.
• May have inaccurate data, make calculation errors or reasoning mistakes.
• Struggles with multiturn conversations and user alignment.

Model Details

The model was created using the following Mergekit YAML configuration:

models:
  - model: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
  - model: Skywork/Skywork-o1-Open-Llama-3.1-8B
  - model: SimpleBerry/LLaMA-O1-Supervised-1129
  - model: NousResearch/DeepHermes-3-Llama-3-8B-Preview
  - model: O1-OPEN/OpenO1-LLama-8B-v0.1
  - model: nvidia/Llama-3.1-Nemotron-Nano-8B-v1
merge_method: karcher
tokenizer:
  source: meta-llama/Llama-3.1-8B-Instruct
dtype: bfloat16

Then the merged model was trained on agentlans/train-of-thought 10K subset for 1 epoch using LLaMA Factory.

• LoRA rank 8, alpha 16, dropout 0.5, use rsLoRA
• Pack sequences, NEFTune 5
• Liger kernel accelerator

Licence

Llama 3.1 license