Dataset · SWE-bench
SWE-smith Tasks for Code Agents
SWE-bench/SWE-smith
50,137 tasks from 128 GitHub repositories for training code repair models.
What's inside
A text dataset of software engineering tasks: 50,137 instances from 128 GitHub repositories collected using the SWE-smith tool. The language is English. The card warns that the monorepository is no longer updated, recommending separate language-specific datasets instead.
What you can do with it
- Training code LLMs and SWE agents
- Fine-tuning LoRA for bug fixing
- Testing on SWE-bench-style tasks
- Tasks
- Text generation
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 84.6K
- Likes
- 70
Models trained on it
- SWE-bench/SWE-agent-LM-7BText generation 2.1K
- mradermacher/SWE-agent-LM-7B-i1-GGUF 1.1K
- SWE-bench/SWE-agent-LM-32BText generation 891
- mradermacher/SWE-agent-LM-7B-GGUF 842
- Mungert/SWE-agent-LM-32B-GGUFText generation 828
- tensorblock/SWE-bench_SWE-agent-LM-32B-GGUFText generation 49
- jasonzhanglll/qwen3.6-35b-a3b-swe-lora-sft-v01Text generation 37
- mlx-community/SWE-agent-LM-32B-4bitText generation 32
How to load it
from datasets import load_dataset
ds = load_dataset("SWE-bench/SWE-smith")