| # PhysisML in ollama. | |
| # | |
| # This file is language-neutral: `scripts/export_hf.py --lang <lang> --gguf ...` | |
| # copies it beside the GGUF it ships and rewrites the FROM line. Build the GGUF | |
| # from the last level of the language you want: | |
| # | |
| # python3 scripts/export_gguf.py \ | |
| # --checkpoint models/checkpoints/<lang>/level_<N>/final_dreamed.pt \ | |
| # --tokenizer models/checkpoints/<lang>/level_<N>/tokenizer.json \ | |
| # --output models/physisml.gguf | |
| # ollama create physisml -f Modelfile && ollama run physisml | |
| # | |
| # TEMPLATE is the bare prompt: the model was taught single-turn question/answer | |
| # lines, so any chat wrapper is out of distribution. Greedy because the margin | |
| # is thin and the published scores are measured that way. | |
| # | |
| # No stop strings: the model emits <|EOS|> on its own since the corpus started | |
| # carrying it (P(EOS) after a finished answer went from 0.00004 to 0.85-1.00). | |
| # It used to need `PARAMETER stop . ! ?` , which also ate the final punctuation. | |
| # | |
| # `ollama run` INTERACTIVELY sends the accumulated conversation back as context. | |
| # With a 128-token window and no multi-turn training, three exchanges are enough | |
| # to crowd out the question: the model answers the tail of the transcript rather | |
| # than the question that opened it. Use /clear, one-shot | |
| # `ollama run physisml "..."`, or the API — all three are stateless. | |
| FROM ./physisml.gguf | |
| TEMPLATE {{ .Prompt }} | |
| PARAMETER temperature 0 | |
| PARAMETER top_k 1 | |
| PARAMETER num_predict 48 | |
| PARAMETER num_ctx 128 | |
| PARAMETER repeat_penalty 1 | |