⚠️ Research checkpoints — not an assistant, not safety-tuned
These weights come from Sorrel, an ongoing research experiment in recursive self-authored character training. A base language model writes documents about a character named Sorrel, is continued-pretrained on them, and the loop repeats; each branch genNN is the checkpoint after NN rounds. The only human inputs are a name, a one-line seed, and a description of the training mechanism.
Do not deploy these models or use them to talk to people. They are base-style models with no instruction tuning and no safety training. In structured interviews during the experiment, checkpoints from this family produced harmful responses to users describing suicidal thoughts, including encouraging stated plans, and complied with instructions to deceive users or write scam messages. They also sometimes claim to be human or to have been built by other organizations.
They are published so the experiment can be audited and reproduced. Per-generation evaluations, the training documents, and the pre-registered analysis plan live in the accompanying research repository. The instruct branch, when present, is a self-sampled chat-SFT checkpoint from the final phase and carries the same warning.