README.md

---
license: unknown
library_name: peft
tags:
- llama-2
datasets:
- ehartford/dolphin
- garage-bAInd/Open-Platypus
inference: false
pipeline_tag: text-generation
base_model: meta-llama/Llama-2-7b-hf
---

<div align="center">

<img src="./assets/llama.png" width="150px">

</div>

# Llama-2-7B-Instruct-v0.1

This instruction model was built via parameter-efficient QLoRA finetuning of [Llama-2-7b](https://huggingface.co/meta-llama/Llama-2-7b-hf) on the first 5k rows of [ehartford/dolphin](https://huggingface.co/datasets/ehartford/dolphin) and the first 5k rows of [garage-bAInd/Open-Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus). Finetuning was executed on 1x A100 (40 GB SXM) for roughly 2 hours on the [Lambda Labs](https://cloud.lambdalabs.com/instances) platform.

## Benchmark metrics

| Metric                | Value |
|-----------------------|-------|
| MMLU (5-shot)         | 46.63 |
| ARC (25-shot)         | 51.19 |
| HellaSwag (10-shot)   | 78.92 |
| TruthfulQA (0-shot)   | 48.5 |
| Avg.                  | 56.31 |

We use the [Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) to run the benchmark tests above, using the same version as Hugging Face's [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard).

## Helpful links

* Model license: coming
* Basic usage: coming
* Finetuning code: coming
* Loss curves: coming
* Runtime stats: coming

## Loss curve

![loss curve](https://raw.githubusercontent.com/daniel-furman/sft-demos/main/assets/sep_12_23_9_20_00_log_loss_curves_Llama-2-7b-instruct.png)

The above loss curve was generated from the run's private wandb.ai log.  

## Limitations and biases

_The following language is modified from [EleutherAI's GPT-NeoX-20B](https://huggingface.co/EleutherAI/gpt-neox-20b)_

This model can produce factually incorrect output, and should not be relied on to produce factually accurate information.
This model was trained on various public datasets.
While great efforts have been taken to clean the pretraining data, it is possible that this model could generate lewd, biased or otherwise offensive outputs.

## How to use

* [notebook](assets/basic_inference_llama_2_dolphin.ipynb)

```python
!pip install -q -U huggingface_hub peft transformers torch accelerate
```

```python
from huggingface_hub import notebook_login
import torch
from peft import PeftModel, PeftConfig
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
    pipeline,
)

notebook_login()
```

```python
peft_model_id = "dfurman/Llama-2-7B-Instruct-v0.1"
config = PeftConfig.from_pretrained(peft_model_id)

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    quantization_config=bnb_config,
    use_auth_token=True,
    device_map="auto",
)

tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path, use_fast=True)
tokenizer.pad_token = tokenizer.eos_token

model = PeftModel.from_pretrained(model, peft_model_id)

format_template = "You are a helpful assistant. {query}\n"
```

```python
# First, format the prompt
query = "Tell me a recipe for vegan banana bread."
prompt = format_template.format(query=query)

# Inference can be done using model.generate
print("\n\n*** Generate:")

input_ids = tokenizer(prompt, return_tensors="pt").input_ids.cuda()
with torch.autocast("cuda", dtype=torch.bfloat16):
    output = model.generate(
        input_ids=input_ids,
        max_new_tokens=512,
        do_sample=True,
        temperature=0.7,
        return_dict_in_generate=True,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
        repetition_penalty=1.2,
    )

print(tokenizer.decode(output["sequences"][0], skip_special_tokens=True))
```

## Runtime tests

coming

## Acknowledgements

This model was finetuned by Daniel Furman on Sep 10, 2023 and is for research applications only.

## Disclaimer

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please cosult an attorney before using this model for commercial purposes.

## meta-llama/Llama-2-7b-hf citation

```
coming
```

## Training procedure

The following `bitsandbytes` quantization config was used during training:
- quant_method: bitsandbytes
- 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: False
- bnb_4bit_compute_dtype: bfloat16

## Framework versions

- PEFT 0.6.0.dev0

# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_dfurman__llama-2-7b-instruct-peft)

| Metric                | Value                     |
|-----------------------|---------------------------|
| Avg.                  | 44.5   |
| ARC (25-shot)         | 51.19          |
| HellaSwag (10-shot)   | 78.92    |
| MMLU (5-shot)         | 46.63         |
| TruthfulQA (0-shot)   | 48.5   |
| Winogrande (5-shot)   | 74.43   |
| GSM8K (5-shot)        | 5.99        |
| DROP (3-shot)         | 5.82         |