| # TinyBalls-V1-Base — base (non-instruct) model for Ollama. | |
| # | |
| # There is no TEMPLATE and no SYSTEM here on purpose. This checkpoint never saw | |
| # chat supervision: the ChatML tokens (<|im_start|>, <|im_end|>) are inert in | |
| # it, verified directly — p(im_end) ~1e-8 everywhere, and a two-turn ChatML | |
| # generation answers "The capital of Germany is Paris." and then degenerates | |
| # without ever emitting im_end or eos. Wrapping it in a chat template would | |
| # advertise a capability it does not have. | |
| # | |
| # Use `ollama run` with a raw completion, or point any completion-style client | |
| # at it. Documents are separated by <|endoftext|> (id 0), which is also the EOS | |
| # and the only stop you need. | |
| FROM ./model.safetensors | |
| PARAMETER temperature 0.8 | |
| PARAMETER top_p 0.95 | |
| PARAMETER num_ctx 2048 | |
| PARAMETER num_predict 512 | |