Hunch — working checkpoint backup
This is not a release. It is an off-site copy of training checkpoints that otherwise existed only on rented GPU instances that can be cancelled without notice. It is public because the free tier caps private storage, not because these weights are ready to use.
Each ckpt/<run>/best/ holds model.safetensors and a RECEIPT.json recording its sha256, byte
count and the host it came from, so any copy can be verified against the machine that produced it.
These are flat scorers built on Qwen3 backbones: AutoModel plus an fp32 RMSNorm and a Linear(d, 1)
head, with a per-question softmax over candidates. They are not chat models and will not respond
to prompts. Loading one needs the matching code, and the runs differ in size, data version, seed and
step count — several are mid-training snapshots.
Known limitations of the wider work, carried here so nothing on this page overstates it: the release candidate is calibrated in family but its fitted temperature does not transfer out of family; on indirect-answer interpretation (Circa) it scores below a constant predictor; and the model-selection rule was amended while results were visible. Anyone evaluating these weights should read the model card in the source repository rather than treating this backup as documentation.
Apache-2.0, consistent with the intended release licence.