File size: 11,170 Bytes
ca7a92e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 |
# QLoRA Instruction Tuned Models
| [Paper](https://arxiv.org/abs/2305.14314) | [Code](https://github.com/artidoro/qlora) | [Demo](https://huggingface.co/spaces/uwnlp/guanaco-playground-tgi) |
**The `QLoRA Instruction Tuned Models` are open-source models obtained through 4-bit QLoRA tuning of LLaMA base models on various instruction tuning datasets. They are available in 7B, 13B, 33B, and 65B parameter sizes.**
**Note: The best performing chatbot models are named [Guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco) and finetuned on OASST1. This model card is for the other models finetuned on other instruction tuning datasets.**
⚠️ These models are purely intended for research purposes and could produce problematic outputs.
## What are QLoRA Instruction Tuned Models and why use them?
- **Strong performance on MMLU** following the QLoRA instruction tuning.
- **Replicable and efficient instruction tuning procedure** that can be extended to new use cases. QLoRA training scripts are available in the [QLoRA repo](https://github.com/artidoro/qlora).
- **Rigorous comparison to 16-bit methods** (both 16-bit full-finetuning and LoRA) in [our paper](https://arxiv.org/abs/2305.14314) demonstrates the effectiveness of 4-bit QLoRA finetuning.
- **Lightweight** checkpoints which only contain adapter weights.
## License and Intended Use
QLoRA Instruction Tuned adapter weights are available under Apache 2 license. Note the use of these adapter weights, requires access to the LLaMA model weighs and therefore should be used according to the LLaMA license.
## Usage
Here is an example of how you would load Flan v2 7B in 4-bits:
```python
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
model_name = "huggyllama/llama-7b"
adapters_name = 'timdettmers/qlora-flan-7b'
model = AutoModelForCausalLM.from_pretrained(
model_name,
load_in_4bit=True,
torch_dtype=torch.bfloat16,
device_map="auto",
max_memory= {i: '24000MB' for i in range(torch.cuda.device_count())},
quantization_config=BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type='nf4'
),
)
model = PeftModel.from_pretrained(model, adapters_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
```
Inference can then be performed as usual with HF models as follows:
```python
prompt = "Introduce yourself"
formatted_prompt = (
f"A chat between a curious human and an artificial intelligence assistant."
f"The assistant gives helpful, detailed, and polite answers to the user's questions.\n"
f"### Human: {prompt} ### Assistant:"
)
inputs = tokenizer(formatted_prompt, return_tensors="pt").to("cuda:0")
outputs = model.generate(inputs=inputs.input_ids, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```
Expected output similar to the following:
```
A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.
### Human: Introduce yourself ### Assistant: I am an artificial intelligence assistant. I am here to help you with any questions you may have.
```
## Current Inference Limitations
Currently, 4-bit inference is slow. We recommend loading in 16 bits if inference speed is a concern. We are actively working on releasing efficient 4-bit inference kernels.
Below is how you would load the model in 16 bits:
```python
model_name = "huggyllama/llama-7b"
adapters_name = 'timdettmers/qlora-flan-7b'
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
device_map="auto",
max_memory= {i: '24000MB' for i in range(torch.cuda.device_count())},
)
model = PeftModel.from_pretrained(model, adapters_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
```
## Model Card
**Architecture**: The models released here are LoRA adapters to be used on top of LLaMA models. They are added to all layers. For all model sizes, we use $r=64$.
**Base Model**: These models use LLaMA as base model with sizes 7B, 13B, 33B, 65B. LLaMA is a causal language model pretrained on a large corpus of text. See [LLaMA paper](https://arxiv.org/abs/2302.13971) for more details. Note that these models can inherit biases and limitations of the base model.
**Finetuning Data**: These models are finetuned on various instruction tuning datasets. The datasets used are: Alpaca, HH-RLHF, Unnatural Instr., Chip2, Longform, Self-Instruct, FLAN v2.
**Languages**: The different datasets cover different languages. We direct to the various papers and resources describing the datasets for more details.
Next, we describe Training and Evaluation details.
### Training
QLoRA Instruction Tuned Models are the result of 4-bit QLoRA supervised finetuning on different instruction tuning datasets.
All models use NormalFloat4 datatype for the base model and LoRA adapters on all linear layers with BFloat16 as computation datatype. We set LoRA $r=64$, $\alpha=16$. We also use Adam beta2 of 0.999, max grad norm of 0.3 and LoRA dropout of 0.1 for models up to 13B and 0.05 for 33B and 65B models.
For the finetuning process, we use constant learning rate schedule and paged AdamW optimizer.
### Training hyperparameters
| Parameters | Dataset | Batch size | LR | Steps | Source Length | Target Length |
|------------|----------|------------|------|-------|---------------|---------------|
| 7B | All | 16 | 2e-4 | 10000 | 384 | 128 |
| 7B | OASST1 | 16 | 2e-4 | 1875 | - | 512 |
| 7B | HH-RLHF | 16 | 2e-4 | 10000 | - | 768 |
| 7B | Longform | 16 | 2e-4 | 4000 | 512 | 1024 |
| 13B | All | 16 | 2e-4 | 10000 | 384 | 128 |
| 13B | OASST1 | 16 | 2e-4 | 1875 | - | 512 |
| 13B | HH-RLHF | 16 | 2e-4 | 10000 | - | 768 |
| 13B | Longform | 16 | 2e-4 | 4000 | 512 | 1024 |
| 33B | All | 32 | 1e-4 | 5000 | 384 | 128 |
| 33B | OASST1 | 16 | 1e-4 | 1875 | - | 512 |
| 33B | HH-RLHF | 32 | 1e-4 | 5000 | - | 768 |
| 33B | Longform | 32 | 1e-4 | 2343 | 512 | 1024 |
| 65B | All | 64 | 1e-4 | 2500 | 384 | 128 |
| 65B | OASST1 | 16 | 1e-4 | 1875 | - | 512 |
| 65B | HH-RLHF | 64 | 1e-4 | 2500 | - | 768 |
| 65B | Longform | 32 | 1e-4 | 2343 | 512 | 1024 |
### Evaluation
We use the MMLU benchmark to measure performance on a range of language understanding tasks. This is a multiple-choice benchmark covering 57 tasks including elementary mathematics, US history, computer science, law, and more. We report 5-shot test accuracy.
Dataset | 7B | 13B | 33B | 65B
---|---|---|---|---
LLaMA no tuning | 35.1 | 46.9 | 57.8 | 63.4
Self-Instruct | 36.4 | 33.3 | 53.0 | 56.7
Longform | 32.1 | 43.2 | 56.6 | 59.7
Chip2 | 34.5 | 41.6 | 53.6 | 59.8
HH-RLHF | 34.9 | 44.6 | 55.8 | 60.1
Unnatural Instruct | 41.9 | 48.1 | 57.3 | 61.3
OASST1 (Guanaco) | 36.6 | 46.4 | 57.0 | 62.2
Alpaca | 38.8 | 47.8 | 57.3 | 62.5
FLAN v2 | 44.5 | 51.4 | 59.2 | 63.9
We evaluate the generative language capabilities through automated evaluations on the Vicuna benchmark. We report the score of the QLoRA Instruction Finetuned Models relative to the score obtained by ChatGPT. The rater in this case is GPT-4 which is tasked to assign a score out of 10 to both ChatGPT and the model outputs for each prompt. We report scores for models ranging 7B to 65B and compare them to both academic and commercial baselilnes.
| Model / Dataset | Params | Model bits | Memory | ChatGPT vs Sys | Sys vs ChatGPT | Mean | 95\% CI |
|------------------|--------|------------|--------|----------------|----------------|------------------|---------|
| GPT-4 | - | - | - | 119.4\% | 110.1\% | **114.5**\% | 2.6\% |
| Bard | - | - | - | 93.2\% | 96.4\% | 94.8\% | 4.1\% |
| Guanaco | 65B | 4-bit | 41 GB | 96.7\% | 101.9\% | **99.3**\% | 4.4\% |
| Alpaca | 65B | 4-bit | 41 GB | 63.0\% | 77.9\% | 70.7\% | 4.3\% |
| FLAN v2 | 65B | 4-bit | 41 GB | 37.0\% | 59.6\% | 48.4\% | 4.6\% |
| Guanaco | 33B | 4-bit | 21 GB | 96.5\% | 99.2\% | **97.8**\% | 4.4\% |
| Open Assistant | 33B | 16-bit | 66 GB | 73.4\% | 85.7\% | 78.1\% | 5.3\% |
| Alpaca | 33B | 4-bit | 21 GB | 67.2\% | 79.7\% | 73.6\% | 4.2\% |
| FLAN v2 | 33B | 4-bit | 21 GB | 26.3\% | 49.7\% | 38.0\% | 3.9\% |
| Vicuna | 13B | 16-bit | 26 GB | 91.2\% | 98.7\% | **94.9**\% | 4.5\% |
| Guanaco | 13B | 4-bit | 10 GB | 87.3\% | 93.4\% | 90.4\% | 5.2\% |
| Alpaca | 13B | 4-bit | 10 GB | 63.8\% | 76.7\% | 69.4\% | 4.2\% |
| HH-RLHF | 13B | 4-bit | 10 GB | 55.5\% | 69.1\% | 62.5\% | 4.7\% |
| Unnatural Instr. | 13B | 4-bit | 10 GB | 50.6\% | 69.8\% | 60.5\% | 4.2\% |
| Chip2 | 13B | 4-bit | 10 GB | 49.2\% | 69.3\% | 59.5\% | 4.7\% |
| Longform | 13B | 4-bit | 10 GB | 44.9\% | 62.0\% | 53.6\% | 5.2\% |
| Self-Instruct | 13B | 4-bit | 10 GB | 38.0\% | 60.5\% | 49.1\% | 4.6\% |
| FLAN v2 | 13B | 4-bit | 10 GB | 32.4\% | 61.2\% | 47.0\% | 3.6\% |
| Guanaco | 7B | 4-bit | 5 GB | 84.1\% | 89.8\% | **87.0**\% | 5.4\% |
| Alpaca | 7B | 4-bit | 5 GB | 57.3\% | 71.2\% | 64.4\% | 5.0\% |
| FLAN v2 | 7B | 4-bit | 5 GB | 33.3\% | 56.1\% | 44.8\% | 4.0\% |
## Citation
```bibtex
@article{dettmers2023qlora,
title={QLoRA: Efficient Finetuning of Quantized LLMs},
author={Dettmers, Tim and Pagnoni, Artidoro and Holtzman, Ari and Zettlemoyer, Luke},
journal={arXiv preprint arXiv:2305.14314},
year={2023}
}
``` |