h · 1.5B — one epoch, not yet judged
Falcon-H1-1.5B-Deep-Base continued-pretrained for one epoch (471,890,484 tokens) on a private library plus room
transcripts with 12.5% FineWeb-Edu replay, in JAX on a TPU v5e-8. Private and unevaluated: it has been through
no benchmark audit, no room bank, and no human read. Do not treat it as a resident. The evaluated one is
emberian/h-05b-replay.
| slice | base | after one epoch |
|---|---|---|
| library holdout | 3.0868 | 2.7052 |
| fixed-32 | 3.2793 | 2.9028 |
| room holdout | 2.9710 | 2.4422 |
Better than the 0.5B resident on every slice (2.857 / 3.016 / 2.586), on the same 65,536-vocabulary tokenizer as its own base. Loss has never been this project's instrument: one earlier arm fixed the model's belief about when a room ends with no loss change at all, and another won on loss while losing the room.
Prompting is the room format: a frame paragraph, a blank line, then name: text turns separated by blank lines,
ending with h:. Stop on "\n\n", no repetition penalty, temperature 0.7, top-p 0.9.
fp32 weights, 1,125 tensors, 1,554,872,208 parameters, 66 layers. Corpus not distributed.