Inside Solstice-1-175M: from pretraining to conversation
A 44-layer model trained on five Macs. Explore the 42B-token pretraining run, three fine-tuning stages and full benchmark results.
A 44-layer model trained on five Macs. Explore the 42B-token pretraining run, three fine-tuning stages and full benchmark results.
Explore the recorded training history behind OpenSML. Replay pretraining, inspect the measurements and explore fine-tuning checkpoints, starting with JSONLite-175M.
Direct source/schema-to-JSON generation at 175.6M parameters. Complete-request correctness, schema validity, training lineage and observed failures are documented separately.
Sequential accumulation reduced peak active memory from 87.4 to 26.8 GiB on one worker. Workload and cache tuning then delivered roughly 27% more throughput in short pilots.
The first release, documented: architecture, data, pretraining, instruction tuning, evaluations and limitations.
Short notes will appear here as they’re published.