OpenLLMs: Less is More for Open-source Models

OpenLLMs is a series of open-source language models fine-tuned on a small, yet diverse and high-quality dataset of multi-round conversations. Specifically, we utilize only ~6K GPT-4 conversations directly filtered from the ~90K ShareGPT conversations. Despite the small size of the dataset, OpenLLMs has demostrated remarkable performance.

🔥 80.9% win-rate, rank #1 of open-source models on AlpacaEval

🚀 105.7% ChatGPT score on Vicuna GPT-4 eval

🤗 Using only 6K data

News

• [2023/07] We released the OpenLLMs model series. Among them, OpenChat obtains 80.9% win-rate on AlpacaEval and 105% ChatGPT performance on Vicuna GPT-4 evaluation.

Models and Dataset

⚠️ Note: The evaluation metrics represent a quantified measure of a subset of the model's capabilities. A score of 105% does not necessarily indicate that the model is better than ChatGPT in all scenarios or for all use cases. It is important to consider the specific tasks or applications for which the model was evaluated and compare the results accordingly.

Generic Models:

• OpenChat: based on LLaMA-13B with a context length of 2048.
  • Achieves 105.7% of ChatGPT score on the Vicuna GPT-4 evaluation.
  • Achieves 80.9% win-rate on AlpacaEval.
• OpenChat-8192: based on LLaMA-13B, with an extended context length of 8192.
  • Achieves 106.6% of ChatGPT score on the Vicuna GPT-4 evaluation.
  • Achieves 79.5% win-rate on AlpacaEval.

Code Models:

• OpenCoderPlus: based on StarCoderPlus with a native context length of 8192.
  • Achieves 102.5% of ChatGPT score on the Vicuna GPT-4 evaluation.
  • Achieves a 78.7% win-rate on AlpacaEval.

Dataset:

• openchat_sharegpt4_dataset: ~6k cleaned and filtered GPT-4 data from ShareGPT.

Model Evaluation

We have evaluated our models using the two most popular evaluation benchmarks, including Vicuna GPT-4 and AlpacaEval benchmarks. The evaluation results are presented in the following figures.

Vicuna Evaluation

Considering that our fine-tuning dataset is produced by GPT-4, we use both GPT-4 and GPT-3.5-Turbo as evaluators, respectively. Note that our evaluation schema slightly differs from Vicuna's. Following Wang et. al, 2023, we additionally adopted evidence calibration (EC) + balanced position calibration (BPC) to reduce potential bias.

Vicuna GPT-4 Evaluation (v.s. gpt-3.5-turbo)

Vicuna GPT-3.5-Turbo Evaluation (v.s. gpt-3.5-turbo)

AlpacaEval

Here we list the minimal version of AlpacaEval with our released models. The full version of AlpacaEval can be found on this page.

Standard benchmarks (In progress)

Due to the limitations of Vicuna GPT-4 Evaluation and AlpacaEval, we are trying to use extensive standard benchmarks to evaluate the performance of OpenLLMs.

Installation

To use OpenLLMs, you need to have CUDA and PyTorch installed. You can clone this repository and install the dependencies via pip:

git clone git@github.com:imoneoi/OChat.git

pip install --no-build-isolation flash-attn
pip install -r requirements.txt

Weights & Serving

We provide full weights of all models as huggingface repos. You can use the following commands to start a local API server at http://localhost:18888. Please note that models should be used under their foundation models' license.

The server is compatible with the ChatCompletions protocol (please note that some functionalities are not fully supported) and the openai package. You can specify the server of openai package by setting:

openai.api_base = "http://localhost:18888/v1"

The currently supported ChatCompletions arguments are:

We also provide a Web UI for a better user experience, please refer to the following section for details.

Note: We recommend having a GPU with memory of at least 40GB (1x A100) to run the server.

Web UI

You can interact with the model using OpenChat-UI, which is a fork of Chatbot UI with support for OpenChat models.

To use OpenChat-UI, follow these steps:

1. Clone the OpenChat-UI repo:

git clone https://github.com/imoneoi/openchat-ui.git

2. Install Dependencies

npm i

3. Set the API host to the local server (or the address of the OpenChat server)

Create a .env.local file in the root of the OpenChat-UI repo with the following content:

OPENAI_API_HOST=http://localhost:18888
OPENAI_API_KEY=openchat-dummy-key
NEXT_PUBLIC_DEFAULT_TEMPERATURE=0.7

4. Run the App

npm run dev

Model modifications

We added an EOT (end-of-turn) token to every base model. For LLaMA models, the embedding of EOT is initialized as the average of all existing token embeddings. For StarCoder models, the embedding of EOT is randomly initialized with 0.02 standard deviation.

For LLaMA-based models with 8192 context, the max_position_embeddings was set to 8192, and RoPE codes were extrapolated. An attempt to interpolate the RoPE code was made, but it resulted in a significant drop in performance (~101% Vicuna GPT-4 evaluation) without mixing pretraining data.

Dataset

🤗 Converted dataset available at openchat_sharegpt4_dataset

The dataset used in the project is a cleaned and filtered version of ShareGPT, retaining only GPT-4 conversations. The original ShareGPT contained approximately 90K conversations, and only 6K cleaned GPT-4 conversations were retained for fine-tuning.

The cleaned GPT-4 conversations were combined with conversation templates and end-of-turn tokens, then cut to the context limit of the model (further content was simply discarded).

To run the data pipeline, execute the following command:

./ochat/data/run_data_pipeline.sh INPUT_FOLDER OUTPUT_FOLDER

The input folder should contain a ShareGPT folder with .html files for each ShareGPT conversation page inside.

The data pipeline consists of three steps:

• Cleaning: HTML cleaning and conversion to Markdown, removing conversations with the wrong format, removing conversations with blocked words, and hash-based exact deduplication
• Filtering: Preserving only conversations marked as Model: GPT-4
• Converting: Converting and tokenizing all conversations for finetuning

The final converted dataset follows the format:

MODEL_TYPE.train.json / .eval.json

[
    [token_id_list, supervise_mask_list],
    [token_id_list, supervise_mask_list],
    ...
]

MODEL_TYPE.train.text.json / .eval.text.json

Plain text decoded from token_id_list

Dataset visualization

We provide a tool for visualizing the embeddings of conversations. To use this tool, open ochat/visualization/ui/visualizer.html using a browser and drag MODEL_TYPE.visualizer.json into the webpage. Click on 3D plot points to show the corresponding conversation.

The embeddings are created using openai_embeddings.py to calculate embeddings of conversations, then UMAP dimension reduction and K-Means coloring with dim_reduction.ipynb.

Training

The hyperparameters used in training the models are the same across all models:

To train using 8xA100 80GB:

NUM_GPUS=8

deepspeed --num_gpus=$NUM_GPUS --module ochat.training_deepspeed.train \
    --model_type MODEL_TYPE \
    --model_path BASE_MODEL_PATH \
    --save_path TARGET_FOLDER \
    --length_grouping \
    --epochs 5 \
    --data_path DATASET_PATH \
    --deepspeed \
    --deepspeed_config ochat/training_deepspeed/deepspeed_config.json

Evaluation

To run the Vicuna GPT-4 evaluation, follow these steps:

1. Generate model answers

python -m ochat.evaluation.get_model_answer --model_type MODEL_TYPE --models_path PATH_CONTAINING_ALL_MODELS_SAME_TYPE --data_path ./ochat/evaluation/vicuna --output_path ./eval_results

2. Generate baseline (GPT-3.5) answers

OPENAI_API_KEY=sk-XXX python -m ochat.evaluation.get_openai_answer --data_path ./ochat/evaluation/vicuna --output_path ./eval_baselines --model_types gpt-3.5-turbo

3. Run GPT-4 evaluation

OPENAI_API_KEY=sk-XXX python -m ochat.evaluation.openai_eval --data_path ./ochat/evaluation/vicuna --baseline_path ./eval_baselines/vicuna_gpt-3.5-turbo.jsonl --input_path ./eval_results

4. Visualize and plot

To visualize and plot the evaluation results, use ochat/visualization/eval_result_ui/eval_result_visualizer.html. Open the file using a browser and select all files inside ./eval_results/eval_result_YYYYMMDD to show the results.

Benchmarks

The same routine as ChatGPT / GPT-4 was used to run other benchmarks or evaluations such as AlpacaEval. Simply run the API server and set the openai.api_base of the benchmark program.

Limitations

Foundation Model Limitations Despite its advanced capabilities, OpenLLMs is still bound by the limitations inherent in its foundation models. These limitations may impact the model's performance in areas such as:

• Complex reasoning
• Mathematical and arithmetic tasks
• Programming and coding challenges

Hallucination of Non-existent Information OpenLLMs may sometimes generate information that does not exist or is not accurate, also known as "hallucination". Users should be aware of this possibility and verify any critical information obtained from the model.

TODO

• Improving conversation splitting
• Mixing SFT data with pretraining data (e.g. RedPajama)
• Extending context by interpolating RoPE (requires mixing with pretraining data)
• Trying LIMA dropout (to determine its usefulness)
• Training larger LLaMA models (needs more computing power)
• Support inference with 2x consumer GPUs

License

Our weight license is subject to their corresponding base model. For example, OpenChat and OpenChat-8192 are the same as the model License of LLaMA for non-commercial use only, while OpenCoderPlus is under the License of StarCoder. Furthermore, we should follow Privacy Practices of ShareGPT. The code is released under Apache License 2.0.

Contact

imonenext@gmail.com

Citation

@software{openllms23,
  title = {{OpenLLMs: Less is More for Open-source Models}},
  author = {Wang, Guan and Cheng, Sijie and Yu, Qiying and Liu, Changling},
  doi = {10.5281/zenodo.8105775},
  url = {https://github.com/imoneoi/openchat},
  version = {pre-release},
  year = {2023},
  month = {7},
}

Acknowledgements

We thank the great work by LLaMA, self-instruct, FastChat (Vicuna), Alpaca and StarCoder.

We also thank the great support by GPT Desk Pte. Ltd. and Tsinghua Laboratory of Brain and Intelligence (THBI).