File size: 3,154 Bytes
47a6bb6
 
b525796
 
 
 
 
 
 
 
 
 
 
47a6bb6
b525796
53c72db
b525796
 
53c72db
b525796
 
53c72db
b525796
 
 
 
 
 
53c72db
 
b525796
53c72db
 
b525796
53c72db
 
 
 
b525796
 
53c72db
b525796
 
 
 
 
4bb822c
b525796
 
 
 
 
 
 
53c72db
b525796
23a4823
b525796
 
 
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
---
license: apache-2.0
datasets:
- ai2lumos/lumos_unified_plan_iterative
language:
- en
tags:
- language-agent
- web-agent
- question-answering
- maths
- reasoning
- planning
---

# 馃獎 Agent Lumos: Unified and Modular Training for Open-Source Language Agents
<p align="center">
  馃寪<a href="https://allenai.github.io/lumos">[Website]</a> &nbsp;
  馃摑<a href="https://arxiv.org/abs/2311.05657">[Paper]</a> &nbsp;
  馃<a href="https://huggingface.co/datasets?sort=trending&search=ai2lumos">[Data]</a> &nbsp;
  馃<a href="https://huggingface.co/models?sort=trending&search=ai2lumos">[Model]</a> &nbsp;
  馃<a href="https://huggingface.co/spaces/ai2lumos/lumos_data_demo">[Demo]</a> &nbsp;
</p>

We introduce 馃獎**Lumos**, Language Agents with **Unified** Formats, **Modular** Design, and **Open-Source** LLMs. **Lumos** unifies a suite of complex interactive tasks and achieves competitive performance with GPT-4/3.5-based and larger open-source agents. 

**Lumos** has following features:
* 馃З **Modular Architecture**:
  - 馃З **Lumos** consists of planning, grounding, and execution modules built based on LLAMA-2-7B/13B and off-the-shelf APIs.
  - 馃 **Lumos** utilizes a unified data format that encompasses multiple task types, thereby enabling the developed agent framework to conveniently support a range of interactive tasks.
* 馃實 **Diverse Training Data**:
  - 馃實 **Lumos** is trained with ~56K diverse high-quality subgoal/action annotations from ground-truth reasoning steps in existing benchmarks with GPT-4.
  - 鈿掞笍 **Lumos** data can be instrumental for future research in developing open-source agents for complex interactive tasks.
* 馃殌 **Competitive Performance**:
  - 馃殌 **Lumos** is comparable or even beats **GPT-series** agents on web/complex QA tasks Mind2Web and HotpotQA, and **larger open agents** on math and multimodal tasks.
  - 馃殌 **Lumos** exceeds contemporaneous agents that have been **fine-tuned** with in-domain HotpotQA, Mind2Web and ScienceQA annotations, such as **FiReAct**, **AgentLM**, and **AutoAct**.
  - 馃殌 **Lumos** performs better than open agent baseline formulations including **chain-of-thoughts** and **integrated** training.
  - 馃殌 **Lumos** surpasses larger open LLM agents and domain-specific agents on unseen tasks, WebShop and InterCode_SQL.

## Model Overview
`lumos_unified_plan_iterative` is a **planning** module checkpoint finetuned on **complex QA**, **web agent**, **multimodal** and **maths** tasks in **Lumos-Iterative (Lumos-I)** formulation. 

The training annotation is shown below:

| Training Data | Number |
|---|---|
|[`lumos_unified_plan_iterative`](https://huggingface.co/datasets/ai2lumos/lumos_unified_plan_iterative)|55382|


## Citation

If you find this work is relevant with your research, please feel free to cite our work!
```
@article{yin2023lumos,
  title={Agent Lumos: Unified and Modular Training for Open-Source Language Agents},
  author={Yin, Da and Brahman, Faeze and Ravichander, Abhilasha and Chandu, Khyathi and Chang, Kai-Wei and Choi, Yejin and Lin, Bill Yuchen},
  journal={arXiv preprint arXiv:2311.05657},
  year={2023}
}
```