Roadmap: v1.3 long-term support

Jeff v1.2 and its nine adapters are a community preview: try them and tell us what works and what doesn't, on GitHub issues. In parallel we are building v1.3, a stable base model with long-term support. It is due in about 36 hours, and the official adapters, retrained on it, will follow shortly after.

What v1.3 changes

• The prompt order. The parts of a request that stay the same come first and the changing field comes last, so a server can prepare the unchanging part once and reuse it. That makes serving faster.
• Nothing else. v1.3 keeps the v1.2 training recipe; only the prompt order changes. (A distilled 0.8B base was tested and made no difference.)

What stays fixed for the life of v1.3

• The request format: state, questions and instructions.
• The option rules: named options, keys never bare numbers.
• The answer format: a probability for every option.

What it means for adapters

• Adapters are tied to the exact base they were trained on. Each adapter records its base, and the server will refuse an adapter trained on a different one. A v1.2 adapter will not load on v1.3; it has to be retrained.
• The data format does not change. A data set written to the data guidelines works for v1.2 today and can be retrained on v1.3 unchanged. That is why the guidelines require the changing state field last.
• An official replay data set ships with v1.3: a licence-cleared sample of the base model's training data, to mix into adapter training so an adapter keeps the base's general skills.
• The official adapters will be retrained on v1.3 and replace the v1.2 ones on this site.

If you are building an adapter now, put your effort into the data set. The v1.2 weights you train today are for testing; your data is what you will keep.