iamplus/Llama-2-70b-hf-ChatOrca-v1

Description :

This model is trained on a mix of Orca data and Open Source + Closed Multi-turn Conversation data to create a better reasoning model which is capable of holding multi-turn conversations as well.

The Dataset split description, Prompt description as well as Training Parameters are given below.

Prompt Description :

The prompt template for the first turn looks like this:

<s>[INST] <<SYS>>
{{ system_prompt }}
<</SYS>>

{{ user_message }} [/INST]

The prompt template for the multi-turn conversation looks like this:

<s>[INST] <<SYS>>
{{ system_prompt }}
<</SYS>>

{{ user_msg_1 }} [/INST] {{ model_answer_1 }} </s><s>[INST] {{ user_msg_2 }} [/INST]

This model follows the official Meta's chat model Prompt format. Please refer here : https://huggingface.co/blog/llama2#how-to-prompt-llama-2 on how to prompt the model for single/multi-turn conversations.

Base model : meta-llama/Llama-2-70b-hf

Data :

1. 1M Orca dara (Gpt-4 Orca data - OpenOrca)
2. 1.7M chat data (includes OpenAssistant Chat data, Ultrachat, and many more open source Chat Datasets)
3. 30k OpenPlatypus data

Training Params :

Number of Epochs : 1
Batch Size : 64
Sequence Length : 4096
Learning Rate : 2e-5 (Cosine)
Weight Decay : 0.1
Gradient Clipping : 1.0
Gamma : 0.85
beta_1 : 0.9
beta_2 : 0.95
eps : 1e-5
Precision : bf16
Optimizer : Any Precision AdamW Optimizer