Step up your LLM alignment with Xwin-LM!
Xwin-LM aims to develop and open-source alignment technologies for large language models, including supervised fine-tuning (SFT), reward models (RM), reject sampling, reinforcement learning from human feedback (RLHF), etc. Our first release, built-upon on the Llama2 base models, ranked TOP-1 on AlpacaEval. Notably, it's the first to surpass GPT-4 on this benchmark. The project will be continuously updated.
- 💥 [Nov, 2023] We released XwinCoder-7B, XwinCoder-13B, XwinCoder-34B. Our XwinCoder-34B reached 74.2 on HumanEval and it achieves comparable performance as GPT-3.5-turbo on 6 benchmarks.
- 💥 [Oct 12, 2023] Xwin-LM-7B-V0.2 and Xwin-LM-13B-V0.2 have been released, with improved comparison data and RL training (i.e., PPO). Their winrates v.s. GPT-4 have increased significantly, reaching 59.83% (7B model) and 70.36% (13B model) respectively. The 70B model will be released soon.
- 💥 [Sep, 2023] We released Xwin-LM-70B-V0.1, which has achieved a win-rate against Davinci-003 of 95.57% on AlpacaEval benchmark, ranking as TOP-1 on AlpacaEval. It was the FIRST model surpassing GPT-4 on AlpacaEval. Also note its winrate v.s. GPT-4 is 60.61.
- 🔍 [Sep, 2023] RLHF plays crucial role in the strong performance of Xwin-LM-V0.1 release!
- 💥 [Sep, 2023] We released Xwin-LM-13B-V0.1, which has achieved 91.76% win-rate on AlpacaEval, ranking as top-1 among all 13B models.
- 💥 [Sep, 2023] We released Xwin-LM-7B-V0.1, which has achieved 87.82% win-rate on AlpacaEval, ranking as top-1 among all 7B models.
| Model | Checkpoint | Report | License |
|---|---|---|---|
| Xwin-LM-7B-V0.2 | 🤗 HF Link | 📃Coming soon (Stay tuned) | Llama 2 License |
| Xwin-LM-13B-V0.2 | 🤗 HF Link | Llama 2 License | |
| Xwin-LM-7B-V0.1 | 🤗 HF Link | Llama 2 License | |
| Xwin-LM-13B-V0.1 | 🤗 HF Link | Llama 2 License | |
| Xwin-LM-70B-V0.1 | 🤗 HF Link | Llama 2 License | |
| Xwin-Coder-7B | 🤗 HF Link | 📃Brief introduction | Llama 2 License |
| Xwin-Coder-13B | 🤗 HF Link | Llama 2 License | |
| Xwin-Coder-34B | 🤗 HF Link | Llama 2 License |
Xwin-LM performance on AlpacaEval.
The table below displays the performance of Xwin-LM on AlpacaEval, where evaluates its win-rate against Text-Davinci-003 across 805 questions. To provide a comprehensive evaluation, we present, for the first time, the win-rate against ChatGPT and GPT-4 as well. Our Xwin-LM model family establish a new state-of-the-art performance across all metrics. Notably, Xwin-LM-70B-V0.1 has eclipsed GPT-4 for the first time, achieving an impressive win-rate of 95.57% to Text-Davinci-003 and 60.61% to GPT-4.
| Model | AlpacaEval (winrate %) | AlpacaEval (winrate %) | AlpacaEval (winrate %) |
|---|---|---|---|
| v.s. Text-Davinci-003 | v.s. ChatGPT | v.s. GPT4 | |
| Xwin-LM-70B-V0.1 | 95.57 | 87.50 | 60.61 |
| GPT-4 | 95.28 | 84.66 | 50.00 |
| Xwin-LM-13B-V0.2 | 93.22 | 87.14 | 70.36 |
| WizardLM-70B-V1.0 | 92.91 | 80.19 | 46.70 |
| Llama-2-70B-Chat | 92.66 | 81.97 | 51.19 |
| Xwin-LM-13B-V0.1 | 91.76 | 81.79 | 55.30 |
| ChatGPT | 89.37 | 50.00 | 16.60 |
| Xwin-LM-7B-V0.2 | 89.31 | 79.60 | 59.83 |
| WizardLM-13B-V1.2 | 89.17 | 75.75 | 41.91 |
| Xwin-LM-7B-V0.1 | 87.35 | 76.40 | 47.57 |
| Llama-2-13B-Chat | 81.09 | 64.22 | 30.92 |
Case study on AlpacaEval
| Prompts | GPT4 response | Xwin-LM response |
|---|---|---|
| I've put a plate on top of an apple, and take the plate from kitchen to dinning room. Where is the apple now | The apple is now in the dining room, as it was underneath the plate when you moved the plate from the kitchen to the dining room. | The apple is still in the kitchen, as you have only moved the plate. The apple is under the plate, so if you want to bring the apple to the dining room, you should either move the apple directly or put the apple on the plate and then move the plate to the dining room. |
| Answer the following question. How do you say "good evening" in French. |
In French, "good evening" is said as "bonsoir." | In French, "good evening" is translated to "bonsoir." This greeting is used in the evening and nighttime hours, typically after 6 p.m. It is a combination of "bon," which means "good," and "soir," which means "evening." To pronounce it, say "bone-swahr." |
The following table provides a comparison of Xwin-LMs with other LLMs on NLP foundation tasks in Open LLM Leaderboard.
| Model | MMLU 5-shot | ARC 25-shot | TruthfulQA 0-shot | HellaSwag 10-shot | Average |
|---|---|---|---|---|---|
| Text-davinci-003 | 56.9 | 85.2 | 59.3 | 82.2 | 70.9 |
| Vicuna-13b 1.1 | 51.3 | 53.0 | 51.8 | 80.1 | 59.1 |
| Guanaco 30B | 57.6 | 63.7 | 50.7 | 85.1 | 64.3 |
| WizardLM-7B 1.0 | 42.7 | 51.6 | 44.7 | 77.7 | 54.2 |
| WizardLM-13B 1.0 | 52.3 | 57.2 | 50.5 | 81.0 | 60.2 |
| WizardLM-30B 1.0 | 58.8 | 62.5 | 52.4 | 83.3 | 64.2 |
| Llama-2-7B-Chat | 48.3 | 52.9 | 45.6 | 78.6 | 56.4 |
| Llama-2-13B-Chat | 54.6 | 59.0 | 44.1 | 81.9 | 59.9 |
| Llama-2-70B-Chat | 63.9 | 64.6 | 52.8 | 85.9 | 66.8 |
| Xwin-LM-7B-V0.1 | 49.7 | 56.2 | 48.1 | 79.5 | 58.4 |
| Xwin-LM-13B-V0.1 | 56.6 | 62.4 | 45.5 | 83.0 | 61.9 |
| Xwin-LM-70B-V0.1 | 69.6 | 70.5 | 60.1 | 87.1 | 71.8 |
| Xwin-LM-7B-V0.2 | 50.0 | 56.4 | 49.5 | 78.9 | 58.7 |
| Xwin-LM-13B-V0.2 | 56.6 | 61.5 | 43.8 | 82.9 | 61.2 |
To obtain desired results, please strictly follow the conversation templates when utilizing our model for inference. Our model adopts the prompt format established by Vicuna and is equipped to support multi-turn conversations.
A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: Hi! ASSISTANT: Hello.</s>USER: Who are you? ASSISTANT: I am Xwin-LM.</s>......
from transformers import AutoTokenizer, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Xwin-LM/Xwin-LM-7B-V0.1")
tokenizer = AutoTokenizer.from_pretrained("Xwin-LM/Xwin-LM-7B-V0.1")
(
prompt := "A chat between a curious user and an artificial intelligence assistant. "
"The assistant gives helpful, detailed, and polite answers to the user's questions. "
"USER: Hello, can you help me? "
"ASSISTANT:"
)
inputs = tokenizer(prompt, return_tensors="pt")
samples = model.generate(**inputs, max_new_tokens=4096, temperature=0.7)
output = tokenizer.decode(samples[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(output)
# Of course! I'm here to help. Please feel free to ask your question or describe the issue you're having, and I'll do my best to assist you.Because Xwin-LM is based on Llama2, it also offers support for rapid inference using vLLM. Please refer to vLLM for detailed installation instructions.
from vllm import LLM, SamplingParams
(
prompt := "A chat between a curious user and an artificial intelligence assistant. "
"The assistant gives helpful, detailed, and polite answers to the user's questions. "
"USER: Hello, can you help me? "
"ASSISTANT:"
)
sampling_params = SamplingParams(temperature=0.7, max_tokens=4096)
llm = LLM(model="Xwin-LM/Xwin-LM-7B-V0.1")
outputs = llm.generate([prompt,], sampling_params)
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(generated_text)- Release the source code
- Release more capabilities, such as math, reasoning, and etc.
Please consider citing our work if you use the data or code in this repo.
@software{xwin-lm,
title = {Xwin-LM},
author = {Xwin-LM Team},
url = {https://github.com/Xwin-LM/Xwin-LM},
version = {pre-release},
year = {2023},
month = {9},
}
Thanks to Llama 2, FastChat, AlpacaFarm, and vLLM.