Chat Vector
CHAT VECTOR: A SIMPLE APPROACH TO EQUIP LLMS WITH NEW LANGUAGE CHAT CAPABILITIES https://arxiv.org/pdf/2310.04799.pdf
With the advancements in conversational AI, such as ChatGPT, this paper focuses on exploring developing Large Language Models (LLMs) for non-English languages, especially emphasizing alignment with human preferences. We introduce a computationally efficient method, leveraging βchat vector,β to synergize pre-existing knowledge and behaviors in LLMs, restructuring the conventional training paradigm from continual pretrain SFT RLHF to continual pretrain + chat. Our empirical studies, primarily focused on Traditional Chinese, employ LLaMA2 as the base model and acquire the chat vector by subtracting the pre-trained weights, LLaMA2, from the weights of LLaMA2-chat. Evaluating from three distinct facets, which are toxicity, ability of instruction following and multi-turn dialogue demonstrates the chat vector's superior efficacy in βchattingβ. To confirm the adaptability of our approach, we extend our experiments to include models pre-trained in both Korean and Simplified Chinese, illustrating the versatility of our methodology. Overall, we present a significant solution in aligning LLMs with human preferences efficiently across various languages, accomplished by the chat vector.
Merged LM
- mistral 7b
- chat vector
- neural-chat
- marconroni
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
Metric | Value |
---|---|
Avg. | 71.27 |
AI2 Reasoning Challenge (25-Shot) | 69.20 |
HellaSwag (10-Shot) | 86.26 |
MMLU (5-Shot) | 65.07 |
TruthfulQA (0-shot) | 60.03 |
Winogrande (5-shot) | 80.90 |
GSM8k (5-shot) | 66.19 |
- Downloads last month
- 702
Spaces using aqweteddy/mistral_tv-neural-marconroni 14
Evaluation results
- normalized accuracy on AI2 Reasoning Challenge (25-Shot)test set Open LLM Leaderboard69.200
- normalized accuracy on HellaSwag (10-Shot)validation set Open LLM Leaderboard86.260
- accuracy on MMLU (5-Shot)test set Open LLM Leaderboard65.070
- mc2 on TruthfulQA (0-shot)validation set Open LLM Leaderboard60.030
- accuracy on Winogrande (5-shot)validation set Open LLM Leaderboard80.900
- accuracy on GSM8k (5-shot)test set Open LLM Leaderboard66.190