I finetuned 1.5B Qwen to near GPT-4o level bash generation perf

About I finetuned 1.5B Qwen to near GPT-4o level bash generation perf

EasyCommand: open models, training data, and lessons from teaching small models to write shell commands.

In the maker’s words

Purely a hobby side project to see how far I can push a really small model, using (mostly) automated training pipelines Original mention: https://news.ycombinator.com/item?id=49869735 There were a bunch of requests to release it. Full synthetic data: https://huggingface.co/datasets/dirac-run/ec-training-data Models: https://huggingface.co/dirac-run/ec-1.5b-gguf and https://huggingface.co/dirac-run/ec-0.6b-gguf Cli https://github.com/dirac-run/ec feel free to train/use the data as you wish.

Where people found it

• Hacker NewsShow HN: I finetuned 1.5B Qwen to near GPT-4o level bash generation perf5 points0 comments5 hours ago
• Hacker NewsToken Arbitrage: Use Sol for coding, Luna for compaction, save 84%3 points2 comments2026-08-03
• Hacker NewsGPT-5.6-terra used 48.5% more context than Mimo-2.5-pro3 points1 comments2026-07-15
• Hacker NewsCache hit rate dropping by 20% doubles your agent's bills2 points0 comments2026-07-08
• Hacker NewsDon't Conflate 'Minimal' with Minimal Effort2 points0 comments2026-06-20
• Hacker NewsOSS models decisively overtook Proprietary models in openrouter market share4 points1 comments2026-06-18
• Hacker NewsCache hit rates of Inference are more meaningful than the headline costs11 points2 comments2026-05-31
• Hacker NewsCache hit rates of Inference are more meaningful than the headline costs2 points0 comments2026-05-23
• Hacker NewsPeak lazy engineering: Let the code answer its own questions2 points0 comments2026-05-05
• Hacker NewsHash anchors and Myers diff and single-token anchors: 60% cheaper AI code edits5 points1 comments2026-04-26