Research log
-
This notebook goes up
sultaan.io starts publishing engineering notes, because the repository stays private.
-
Portfolio review
Sultaan is confirmed as a research track, not a launch dependency. Ten engineering sessions are planned, and LoRA fine-tuning of an open model is named the practical route to a sovereign model.
-
First contact: a new tier, a supervisor and a recorded run
ADR 0026 adds a roughly 33M first-contact tier, a streamed multi-source mixture, a hardware preflight, a streaming tokenizer trainer, a crash-safe overnight supervisor and a local text-corpus path. The same day, a 33.9M-parameter model trained on a 6 GB laptop GPU: loss fell from 9.78 to 3.89 over 20,000 steps on 500 Project Gutenberg books. That run is recorded only on an unmerged branch, and its checkpoint has not been located since.
-
Training-run hardening
A training-health sentinel, atomic and promoted checkpoints, the tier gate, the provenance schema and the mixture bandit; trust tiers for the knowledge graph, typed memory facts and an agent circuit breaker (ADRs 0021 to 0025).
-
Phase 3 begins
The pretraining stack is fixed: a from-scratch Llama-style decoder, FineWeb-Edu streaming, FSDP chosen for future multi-GPU tiers, bf16 and a cosine schedule (ADR 0016).
-
Evaluation suite expanded
A TruthfulQA multiple-choice subset and a custom hallucination probe set join perplexity and MMLU (ADR 0013).
datesEntries before October use the dates of commits and decision records in the repository.