An original decision-model research project: architecture, training, evaluation, weights and runtime.
The product goal is a model designed here and trained from random initialization on traceable datasets. It encodes shared state once, scores dynamic options without autoregressive decoding and returns typed calibrated decisions.
Third-party models such as JEv, TinyJev and Qwen are local teachers, references and data tools. Their code and weights are not part of the model published by this project. The earlier TinyJev fine-tuning path is preserved as a licensed fallback experiment, not presented as our model.
Current work starts with a small encoder and pointer head trained entirely here, then scales only when measured capacity demands it.
.venv-train/bin/python -m training.python.own_train \
--episodes path/to/train.jsonl \
--out artifacts/checkpoints/own/run-001See project lineage and the active initiative.