n33
Adapters from a controlled behavioural experiment on SN97 Albedo, 2026-09-10.
Initialised from the reigning king and trained for 3 epochs at lr 3e-4 on a 1,840-row subset,
with a checkpoint saved each epoch: epoch1/, epoch2/, epoch3/, plus final/.
These are diagnostic probes, not competition candidates. They were trained at deliberately high strength on a narrowly filtered subset in order to test whether supervised fine-tuning actually changes rollout behaviour. They are expected to be worse than the base model on the task, and epoch 3 produced one degenerate (looped) trajectory in 198. Not submitted anywhere.
- Downloads last month
- -
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for sin333/adapter-n33
Base model
wondercat/albedo-qwen3.6-35b-wonder-e3