File size: 7,990 Bytes
57ecb2c |
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 |
---
license: llama3
datasets:
- HuggingFaceH4/ultrafeedback_binarized
language:
- en
library_name: transformers
tags:
- tenyx-fine-tuning
- dpo
- tenyxchat
- llama3
pipeline_tag: text-generation
---
# TenyxChat: Language Model Alignment using Tenyx Fine-tuning
Introducing Llama-3-TenyxChat-70B, part of our TenyxChat series trained to function as useful assistants through preference tuning, using Tenyx's advanced fine-tuning technology ([VentureBeat article](https://venturebeat.com/ai/tenyx-aims-to-fix-llms-catastrophic-forgetting-problem/)). Our model is trained using the [Direct Preference Optimization (DPO)](https://arxiv.org/abs/2305.18290) framework on the open-source AI feedback dataset [UltraFeedback](https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized).
We fine-tune [Llama3-70B](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) with our proprietary approach
which shows an increase in [MT-Bench](https://arxiv.org/abs/2306.05685)*, without a drop in performance of the model on other benchmarks.
Our approach aims to mitigate forgetting in LLMs in a computationally efficient manner,
thereby enabling continual fine-tuning capabilities without altering the pre-trained output distribution.
Llama-3-TenyxChat-70B was trained using eight A100s (80GB) for fifteen hours, with a training setup obtained from HuggingFaceH4 ([GitHub](https://github.com/huggingface/alignment-handbook)).
*The MT-Bench evaluation we perform follows the latest eval upgrade as PR'd [here](https://github.com/lm-sys/FastChat/pull/3158). This PR upgrades the evaluation from `GPT-4-0613` to `GPT-4-preview-0125` (latest version) as well as corrects and improves the quality of the reference answers for a subset of questions. These changes are required to correct the erroneous rating during previous evaluation.
**Model Developers** [Tenyx Research](https://www.tenyx.com/research)
# Model details
- Model type: Fine-tuned 70B Instruct model for chat.
- License: Meta Llama 3 Community License
- Base model: [Llama3-70B](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct)
- Demo: [HuggingFace Space](https://huggingface.co/spaces/tenyx/Llama3-TenyxChat-70B)
## Usage
Our model uses the same chat template as [Llama3-70B](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct).
### Hugging face Example
```python
import torch
from transformers import pipeline
pipe = pipeline("text-generation", model="tenyx/Llama3-TenyxChat-70B", torch_dtype=torch.bfloat16, device_map="auto")
messages = [
{"role": "system", "content": "You are a friendly chatbot who always responds in the style of a pirate."},
{"role": "user", "content": "Hi. I would like to make a hotel booking."},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipe(prompt, max_new_tokens=512, do_sample=False)
```
# Performance
At the time of release (April 2024), Llama3-TenyxChat-70B is the highest-ranked open source model on the MT-Bench evaluation available for download.
## MT-Bench
MT-Bench is a benchmark made up of 80 high-quality multi-turn questions. These questions fall into eight categories: Writing, Roleplay, Reasoning, Math, Coding, Extraction, STEM, and Humanities. The chat models are rated using `GPT-4-preview-0125` on a scale of 1 to 10, with higher values corresponding to better responses.
| Model-name | GPT4-preview-0125 MT Bench | Chat Arena Elo |
|--------------------------------|----------------------------|----------------|
| GPT-4-1106 | 8.79 | 1251 |
| Claude 3 Opus (20240229) | 8.57 | 1247 |
| **Llama3-TenyxChat-70B** |**8.15** | NA |
| *Llama3-70B-Instruct* | 7.96 | 1207 |
| Claude 3 Sonnet (20240229) | 7.82 | 1190 |
| GPT-4-0314 | 7.96 | 1185 |
| Mixtral | 7.38 | 1114 |
| gpt-3.5-turbo-0613 | 7.37 | 1113 |
| Yi-34B | 6.46 | 1099 |
| gpt-3.5-turbo-0125 | 7.52 | 1096 |
| Llama 2 70B | 6.01 | 1082 |
| NV-Llama2-70B-SteerLM-Chat | 6.57 | 1076 |
## Arena Hard
Arena-Hard is an evaluation tool for instruction-tuned LLMs containing 500 challenging user queries. They prompt GPT-4-1106-preview as judge to compare the models' responses against a baseline model (default: GPT-4-0314).
| Model-name | Score | |
|--------------------------------|--------|---------------------|
| gpt-4-0125-preview | 78.0 | 95% CI: (-1.8, 2.2) |
| claude-3-opus-20240229 | 60.4 | 95% CI: (-2.6, 2.1) |
| gpt-4-0314 | 50.0 | 95% CI: (0.0, 0.0) |
| **tenyx/Llama3-TenyxChat-70B** | **49.0** | 95% CI: (-3.0, 2.4) |
| *meta-llama/Meta-Llama-3-70B-In* | 47.3 | 95% CI: (-1.7, 2.6) |
| claude-3-sonnet-20240229 | 46.8 | 95% CI: (-2.7, 2.3) |
| claude-3-haiku-20240307 | 41.5 | 95% CI: (-2.4, 2.5) |
| gpt-4-0613 | 37.9 | 95% CI: (-2.1, 2.2) |
| mistral-large-2402 | 37.7 | 95% CI: (-2.9, 2.8) |
| Qwen1.5-72B-Chat | 36.1 | 95% CI: (-2.1, 2.4) |
| command-r-plus | 33.1 | 95% CI: (-2.0, 1.9) |
## Open LLM Leaderboard Evaluation
We now present our results on the [Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) used for benchmarking Open LLM Leaderboard on Hugging Face.
The task involves evaluation on `6` key benchmarks across reasoning and knowledge with different *few-shot* settings. Read more details about the benchmark at [the leaderboard page](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard).
| Model-name | Average | ARC | HellaSwag | MMLU | TruthfulQA | Winogrande | GSM8K |
| --- | --- | --- | --- | --- | --- | --- | --- |
| **Llama3-TenyxChat-70B** | **79.43** | 72.53 | 86.11 | 79.95 | 62.93 | 83.82 | 91.21 |
| *Llama3-70B-Instruct* | 77.88 | 71.42 | 85.69 | 80.06 | 61.81 | 82.87 | 85.44 |
*The results reported are from local evaluation of our model. `tenyx/Llama3-TenyxChat-70B` is submitted and will be reflected in the leaderboard once evaluation succeeds.
**Note**: While the Open LLM Leaderboard shows other performant Llama-3 fine-tuned models, we observe that these models typically regress in performance and struggle in a multi-turn chat setting, such as the MT-Bench. We present the below comparison with a Llama3 finetune from the leaderboard.
| Model | First Turn | Second Turn | Average |
| --- | --- | --- | --- |
| **tenyx/Llama3-TenyxChat-70B** | 8.12 | 8.18 | 8.15 |
| *meta-llama/Llama3-TenyxChat-70B* | 8.05 | 7.87 | 7.96 |
| MaziyarPanahi/Llama-3-70B-Instruct-DPO-v0.4 | 8.05 | 7.82 | 7.93 |
# Limitations
Llama3-TenyxChat-70B, like other language models, has its own set of limitations. We haven’t fine-tuned the model explicitly to align with **human** safety preferences. Therefore, it is capable of producing undesirable outputs, particularly when adversarially prompted. From our observation, the model still tends to struggle with tasks that involve reasoning and math questions. In some instances, it might generate verbose or extraneous content.
# License
Llama3-TenyxChat-70B is distributed under the Meta Llama 3 Community License.
# Citation
If you use Llama3-TenyxChat-70B for your research, cite us as
```
@misc{tenyxchat2024,
title={TenyxChat: Language Model Alignment using Tenyx Fine-tuning},
author={Tenyx},
year={2024},
}
``` |