---
library_name: peft
license: llama2
datasets:
- vicgalle/alpaca-gpt4
language:
- en
pipeline_tag: text-generation
tags:
- llama-2
- llama
- instruct
- instruction
---

# Info

Adapter model trained with the  [**QloRA**](https://arxiv.org/abs/2305.14314) technique

* 📜 Model license: [Llama 2 Community License Agreement](https://ai.meta.com/resources/models-and-libraries/llama-downloads/)
* 🏛️ Base Model: [Llama-2-70b-hf](https://huggingface.co/meta-llama/Llama-2-70b-hf)
* 🖥️ Machine: Nvidia A100 (40 GB vRAM)
* 💵 Cost: $3.5
* ⌛ Training Time: 3 hour 22 minutes
* 📊 Dataset Used: [vicgalle/alpaca-gpt4](https://huggingface.co/datasets/vicgalle/alpaca-gpt4)

# Evulation Results ([Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard))

|         | Average | ARC (25-shot) | HellaSwag (10-shot) | MMLU (5-shot) | TruthfulQA (0-shot) |
|---------|---------|---------------|---------------------|---------------|--------------------|
| Scores  | 67.3    | 66.38         | 84.51               | 62.75         | 55.57              |

💪 Our current ranking on [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard) is 13.

# Loss Graph

![](https://i.imgur.com/xPRcRyM.png)


## Training procedure

The following `bitsandbytes` quantization config was used during training:

- load_in_8bit: False
- load_in_4bit: True
- llm_int8_threshold: 6.0
- llm_int8_skip_modules: None
- llm_int8_enable_fp32_cpu_offload: False
- llm_int8_has_fp16_weight: False
- bnb_4bit_quant_type: nf4
- bnb_4bit_use_double_quant: True
- bnb_4bit_compute_dtype: bfloat16