{"name":"opal","repo":"xingjian-bai/opal","url":"https://midorreal.com/project/xingjian-bai-opal","repository":"https://github.com/xingjian-bai/opal","description":"OPAL: a reactive optimizer. A coding agent rewrites the optimizer while your model trains.","facts":{"license":{"value":"MIT","source":"https://github.com/xingjian-bai/opal","checked":"21 Nov 2026"},"language":{"value":"Python","source":"https://github.com/xingjian-bai/opal","checked":"21 Nov 2026"},"homepage":{"value":"https://xingjianbai.com/reactive-optimizers/","source":"https://xingjianbai.com/reactive-optimizers/","checked":"21 Nov 2026"},"last_release":{"value":"v0.3.10, 3 Nov 2026","source":"https://github.com/xingjian-bai/opal/releases","checked":"21 Nov 2026"},"releases":{"value":"6","source":"https://github.com/xingjian-bai/opal/releases","checked":"19 Nov 2026"},"contributors":{"value":"2","source":"https://github.com/xingjian-bai/opal/graphs/contributors","checked":"19 Nov 2026"},"archived":{"value":"no","source":"https://github.com/xingjian-bai/opal","checked":"21 Nov 2026"}},"stars":480.0,"writeup":{"what_it_is":"OPAL is a PyTorch optimizer wrapper. A coding agent (Claude Code) can rewrite the optimizer's update program and an analysis program while a model trains. The wrapped run matches the plain run until the agent changes something.","audience":"PyTorch users who want to train models with an optimizer that an agent adjusts during training.","claims":[{"kind":"specific","claim":"Wrapping an optimizer leaves the run bit-for-bit identical to the plain run until the agent changes something.","excerpt":"Until the agent changes\nsomething, the wrapped run is bit-for-bit the same as the plain one.","status":"not_checked"},{"kind":"specific","claim":"The agent is asked to leave the model, data, loss, batch size, step count and evaluation unchanged.","excerpt":"It is\nasked not to touch your model, data, loss, batch size, number of steps or evaluation.","status":"not_checked"},{"kind":"specific","claim":"A new optimizer program runs on trial for 20 steps and reverts to the previous program if it raises an error.","excerpt":"runs on trial for 20\nsteps. If it raises an error, the run goes back to the previous program.","status":"not_checked"},{"kind":"specific","claim":"Improved final metrics on several example runs, such as FID 23.6 versus 35.7 for SiT-B/2.","excerpt":"| SiT-B/2, ImageNet 256, 400K updates | FID-50K ↓ | 35.7 (ADM evaluator; the SiT paper reports 33.0) | 23.6 |","status":"not_checked"},{"kind":"specific","claim":"Its overhead is roughly fixed per agent decision, so it suits long runs and is costly in short ones.","excerpt":"**Long runs suit it best.** An agent decision takes about the same time in any run, about 40 s on average in ours,","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-09"},"why_now":null,"owner_supplied":null,"score":{"verdict":"mid","hype":52,"reality":59,"gap":-7,"momentum":null,"confidence":85,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-11-03T07:31:48+00:00","text":"Latest release v0.3.10","source":"https://github.com/xingjian-bai/opal/releases"},{"date":"2026-10-22T08:20:56+00:00","text":"First release","source":"https://github.com/xingjian-bai/opal/releases"},{"date":"2026-10-21T14:17:50+00:00","text":"Repository created","source":"https://github.com/xingjian-bai/opal"}],"cite":"opal is a GitHub project at https://github.com/xingjian-bai/opal written mainly in Python, described there as \"OPAL: a reactive optimizer. A coding agent rewrites the optimizer while your model trains\".","attribution":"Data from Mid or Real, https://midorreal.com/project/xingjian-bai-opal, checked 22 November 2026. Attribution with a link is required.","still_real":null}