TimeLlama-13b-chat

TimeLlama

TimeLlama is an instruction-finetuned Llama2 series that improves complex temporal reasoning ability.

Model Details

Model Description

In this work, we introduce the first multi-source dataset for explainable temporal reasoning, called ExpTime. The dataset contains 26k examples derived from temporal knowledge graph datasets. Each example includes a context with multiple events, a future event to predict, and an explanation for the prediction in the form of temporal reasoning over the events.

To generate the dataset, we propose a novel knowledge-graph-instructed-generation strategy. The dataset supports the comprehensive evaluation of large language models on complex temporal reasoning, future event prediction, and explainability.

Based on ExpTime, we develop TimeLlaMA, a series of LLM models fine-tuned for explainable temporal reasoning. TimeLlaMA builds on the foundation LLM LLaMA-2 and utilizes instruction tuning to follow prompts for making explanations.

Model Sources

• Repository: https://github.com/chenhan97/TimeLlama
• Paper: https://arxiv.org/abs/2310.01074

Uses

Direct Use

from transformers import LlamaConfig, LlamaTokenizer, LlamaForCausalLM
# Model names: "chrisyuan45/TimeLlama-7b-chat", "chrisyuan45/TimeLlama-13b-chat"
model = LlamaForCausalLM.from_pretrained(
        model_name,
        return_dict=True,
        load_in_8bit=quantization,
        device_map="auto",
        low_cpu_mem_usage=True)
tokenizer = LlamaTokenizer.from_pretrained(model_name)

Finetune

Please check our repository for the detailed finetuning method.