Update README.md

README.md

@@ -8,7 +8,6 @@ pinned: false
 ---
 Model/Data associated with Paper:
 
-
 <h1 align="center"> Training Software Engineering Agents and Verifiers with SWE-Gym </h1>
 
 <p align="center">
@@ -26,9 +25,83 @@ Model/Data associated with Paper:
   <sub><sup>*</sup>Equal contribution, <sup>^</sup>Equal supervision</sub>
 </p>
 
+<p align="center">
+<a href="assets/paper.pdf">📃 Paper</a>
+•
+<a href="https://huggingface.co/SWE-Gym" >🤗 Data & Models</a>
+</p>
+
+We present **SWE-Gym**, the first environment for training real-world software engineering agents.
+We use it to train strong LM agents that achieve state-of-the-art open results on SWE-Bench, with early, promising scaling characteristics as we increase training and inference-time compute.
+
+<p align="center">
+  <img src="https://github.com/SWE-Gym/SWE-Gym/raw/main/assets/images/teaser.jpg" width="100%" alt="teaser">
+</p>
+
+
+---
+
+Progress in agents for software engineering has been limited by the lack of training environments that both include rigorous verification for reinforcement learning and cover the expansive tasks encountered in real-world repository-level engineering.
+
+We introduce SWE-Gym: An Open Environment for Training Software Engineering Agents & Verifiers.
+Our baselines achieve new open SOTA - 32%/26% on SWE-Bench Verified/Lite, with promising scaling trends.
+
+![SWE-Gym Scaling](https://github.com/SWE-Gym/SWE-Gym/raw/main/assets/images/scaling.jpg)
+*SWE-Gym enables scalable improvements for software engineering agents at both training and inference time. Our current results is primarity bottlenecked by training and inference compute, rather than the size of our environment.*
+
+## SWE-Gym Environment
+
+We create SWE-Gym, the first environment for training SWE agents, with **2.4K real tasks from 11 Python repos** & a Lite split of 234 instances. SWE-Gym combines real-world Python tasks, repository context, executable environments, and test verification to train agents for solving software engineering problems.
+
+![SWE-Gym Repo Distribution](https://github.com/SWE-Gym/SWE-Gym/raw/main/assets/images/swe-gym.jpg)
+
+
+## SWE-Gym trains LMs as agents
+
+When fine-tuned on less than 500 agent-environment interaction trajectories sampled from it from GPT-4o and Claude 3.5 Sonnet, we achieve **+14%** absolute gains on SWE-Bench Verified with an 32B LM-powered OpenHands agent.
+
+![OpenHands Performance diff before and after training](https://github.com/SWE-Gym/SWE-Gym/raw/main/assets/images/oh-agent.jpg)
+
+
+## SWE-Gym enables self-improvement
+
+SWE-Gym is also effective across agent scaffolds. With rejection sampling fine-tuning and MeatlessTools scaffold, our 32B and 7B models achieve 20% and 10% respectively on SWE-Bench Lite through self-improvement.
+
+<p align="center">
+  <img src="https://github.com/SWE-Gym/SWE-Gym/raw/main/assets/images/ml-agent.jpg" width="80%" alt="Moatless self-improvement">
+</p>
+
+
+
+## SWE-Gym enables inference-time scaling 
+
+SWE-Gym enables inference-time scaling through verifiers trained on agent trajectories.  
+These verifiers identify most promising solutions via best-of-n selection, together with our learned agents, they achieve 32%/26% on SWE-Bench Verified/Lite, a new open SoTA.
+
+
+![Inference Time Scaling for Moatless Agent](https://github.com/SWE-Gym/SWE-Gym/raw/main/assets/images/inference-ml.jpg)
+*Inference Time Scaling for Moatless Agent*
+
+![Inference Time Scaling for OpenHands Agent](https://github.com/SWE-Gym/SWE-Gym/raw/main/assets/images/inference-oh.jpg)
+*Inference Time Scaling for OpenHands Agent*
+
+
+## Our baselines on SWE-Gym shows strong scaling trends
+
+Lastly, our ablations reveal strong scaling trends - performance is now bottlenecked by train and inference compute, rather than the size of our dataset. Pushing and improving these scaling trends further is an exciting direction for future work.
+
+![](https://github.com/SWE-Gym/SWE-Gym/raw/main/assets/images/scaling.jpg)
 
-We present SWE-Gym, the first environment for training real-world software engineering agents. We use it to train strong LM agents that achieve state-of-the-art open results on SWE-Bench, with early, promising scaling characteristics as we increase training and inference-time compute.
+## Reproducing Results
 
+See [docs/OpenHands.md](docs/OpenHands.md) and [docs/MoatlessTools.md](docs/MoatlessTools.md) for instructions on reproducing results with our training and inference-time results for OpenHands and MoatlessTools agents.
 
-Get started at project page [github.com/SWE-Gym/SWE-Gym](https://github.com/SWE-Gym/SWE-Gym)
+## 📚 Citation
 
+```bibtex
+@misc{pan2024swegym,
+      title={Training Software Engineering Agents and Verifiers with {SWE-Gym}},
+      author={Pan, Jiayi and Wang, Xingyao and Neubig, Graham and Jaitly, Navdeep and Ji, Heng and Suhr, Alane and Zhang, Yizhe},
+      year={2024},
+}
+```