Model Overview
Model license: Llama-2<br> This model is trained based on TheBloke/Llama-2-13B-Chat-fp16 model that is QLoRA finetuned on Photolens/oasst1-langchain-llama-2-formatted dataset.<br>
Prompt Template: Llama-2
<s>[INST] Prompter Message [/INST] Assistant Message </s>
Intended Use
Dataset that is used to finetune base model is optimized for langchain applications.<br> So this model is intended for a langchain LLM.
Training Details
This model took 2:50:27 to train in QLoRA on a single A100 40gb GPU.<br>
- epochs:
1 - train batch size:
8 - eval batch size:
8 - gradient accumulation steps:
1 - maximum gradient normal:
0.3 - learning rate:
2e-4 - weight decay:
0.001 - optimizer:
paged_adamw_32bit - learning rate schedule:
cosine - warmup ratio (linear):
0.03
Models in this series
| Model | Train time | Size (in params) | Base Model |
|---|---|---|---|
| llama-2-7b-langchain-chat | 1:14:16 | 7 billion | NousResearch/Llama-2-7b-chat-hf |
| llama-2-13b-langchain-chat | 2:50:27 | 13 billion | TheBloke/Llama-2-13B-Chat-fp16 |
| Photolens/OpenOrcaxOpenChat-2-13b-langchain-chat | 2:56:54 | 13 billion | Open-Orca/OpenOrcaxOpenChat-Preview2-13B |
