Causal-Dolphin-Agent-LoRA

<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

Causal-Dolphin-Agent-v0.1

This model is a LoRA fine-tune of CausalLM/7B on Eric's wonderful Dolphin dataset, with THUDM/AgentInstruct mixed in both training runs.

Causal-Dolphin-Agent was trained for 3 epochs on 5 million GPT3.5 augmented FLAN instructions & AgentInstruct dataset in ChatML format. It was then trained a further 3 epochs on 1 million GPT4 augmented FLAN instructions with AgentInstruct shuffled in as well.

It achieves the following results on the evaluation set:

• Loss: 2.8435

Prompt Format

Causal-Dolphin-Agent uses ChatML as the prompt format:

<|im_start|>system
You are Dolphin, a helpful AI assistant.<|im_end|>
<|im_start|>user
If Danny owns a bike, then Edward owns a bike. If Edward owns a bike, then Freddy owns a bike. If Danny owns a bike, which of the following statements must be true? Let's think step by step.

I. Edward owns a bike.
II. Freddy owns a bike.
III. Freddy does not own a bike.

Choose one answer:
I only
II only
III only
I and II only
I and III only
<|im_end|>
<|im_start|>assistant

Training and evaluation data

ehartford/dolphin THUDM/AgentInstruct

Training procedure

Causal-Dolphin-Agent was trained for 3 epochs on 5 million GPT3.5 augmented FLAN instructions & AgentInstruct dataset in ChatML format. It was then trained a further 3 epochs on 1 million GPT4 augmented FLAN instructions with AgentInstruct shuffled in as well.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 6e-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-05
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 3

Training results

Framework versions

• Transformers 4.34.1
• Pytorch 2.1.0+cu121
• Datasets 2.14.6
• Tokenizers 0.14.1