DaYin commited on
Commit
d737fba
β€’
1 Parent(s): 808a11d

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +14 -8
README.md CHANGED
@@ -11,25 +11,30 @@ tags:
11
  - planning
12
  ---
13
 
14
- # πŸͺ„ Lumos: Language Agents with Unified Formats, Modular Design, and Open-Source LLMs
15
  <p align="center">
16
  🌐<a href="https://allenai.github.io/lumos">[Website]</a> &nbsp;
17
- πŸ“<a href="">[Paper]</a> &nbsp;
18
  πŸ€—<a href="https://huggingface.co/datasets?sort=trending&search=ai2lumos">[Data]</a> &nbsp;
19
  πŸ€—<a href="https://huggingface.co/models?sort=trending&search=ai2lumos">[Model]</a> &nbsp;
 
20
  </p>
21
 
22
  We introduce πŸͺ„**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.
23
 
24
  **Lumos** has following features:
25
  * 🧩 **Modular Architecture**:
26
- - **Lumos** consists of planning, grounding, and execution modules built based on LLAMA-2-7B.
 
27
  * 🌍 **Diverse Training Data**:
28
- - **Lumos** is trained with ~40K high-quality annotations from ground-truth reasoning steps in existing benchmarks with GPT-4.
 
29
  * πŸš€ **Competitive Performance**:
30
- - πŸš€ **Lumos** outperforms **GPT-4/3.5-based** agents on complex QA and web agent tasks, and **larger open agents** on maths tasks.
31
- - πŸš€ **Lumos** performs better than open agent baseline formulations including **chain-of-thoughts** and **unmodularized** training.
32
- - πŸš€ **Lumos** surpasses larger open LLM agents and domain-specific agents on an unseen task, WebShop.
 
 
33
 
34
  ## Model Overview
35
  `lumos_web_agent_plan_iterative` is a **planning** module checkpoint finetuned on **web agent** task in **Lumos-Iterative (Lumos-I)** formulation.
@@ -46,8 +51,9 @@ The training annotation is shown below:
46
  If you find this work is relevant with your research, please feel free to cite our work!
47
  ```
48
  @article{yin2023lumos,
49
- title={Lumos: Towards Language Agents that are Unified, Modular, and Open Source},
50
  author={Yin, Da and Brahman, Faeze and Ravichander, Abhilasha and Chandu, Khyathi and Chang, Kai-Wei and Choi, Yejin and Lin, Bill Yuchen},
 
51
  year={2023}
52
  }
53
  ```
 
11
  - planning
12
  ---
13
 
14
+ # πŸͺ„ Agent Lumos: Unified and Modular Training for Open-Source Language Agents
15
  <p align="center">
16
  🌐<a href="https://allenai.github.io/lumos">[Website]</a> &nbsp;
17
+ πŸ“<a href="https://arxiv.org/abs/2311.05657">[Paper]</a> &nbsp;
18
  πŸ€—<a href="https://huggingface.co/datasets?sort=trending&search=ai2lumos">[Data]</a> &nbsp;
19
  πŸ€—<a href="https://huggingface.co/models?sort=trending&search=ai2lumos">[Model]</a> &nbsp;
20
+ πŸ€—<a href="https://huggingface.co/spaces/ai2lumos/lumos_data_demo">[Demo]</a> &nbsp;
21
  </p>
22
 
23
  We introduce πŸͺ„**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.
24
 
25
  **Lumos** has following features:
26
  * 🧩 **Modular Architecture**:
27
+ - 🧩 **Lumos** consists of planning, grounding, and execution modules built based on LLAMA-2-7B/13B and off-the-shelf APIs.
28
+ - πŸ€— **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.
29
  * 🌍 **Diverse Training Data**:
30
+ - 🌍 **Lumos** is trained with ~56K diverse high-quality subgoal/action annotations from ground-truth reasoning steps in existing benchmarks with GPT-4.
31
+ - βš’οΈ **Lumos** data can be instrumental for future research in developing open-source agents for complex interactive tasks.
32
  * πŸš€ **Competitive Performance**:
33
+ - πŸš€ **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.
34
+ - πŸš€ **Lumos** exceeds contemporaneous agents that have been **fine-tuned** with in-domain HotpotQA, Mind2Web and ScienceQA annotations, such as **FiReAct**, **AgentLM**, and **AutoAct**.
35
+ - πŸš€ **Lumos** performs better than open agent baseline formulations including **chain-of-thoughts** and **integrated** training.
36
+ - πŸš€ **Lumos** surpasses larger open LLM agents and domain-specific agents on unseen tasks, WebShop and InterCode_SQL.
37
+
38
 
39
  ## Model Overview
40
  `lumos_web_agent_plan_iterative` is a **planning** module checkpoint finetuned on **web agent** task in **Lumos-Iterative (Lumos-I)** formulation.
 
51
  If you find this work is relevant with your research, please feel free to cite our work!
52
  ```
53
  @article{yin2023lumos,
54
+ title={Agent Lumos: Unified and Modular Training for Open-Source Language Agents},
55
  author={Yin, Da and Brahman, Faeze and Ravichander, Abhilasha and Chandu, Khyathi and Chang, Kai-Wei and Choi, Yejin and Lin, Bill Yuchen},
56
+ journal={arXiv preprint arXiv:2311.05657},
57
  year={2023}
58
  }
59
  ```