Trainfer
Trainfer is the current local daemon for shared-weight inference and incremental learning, descended from Lile in the AGI project.
Project status: Current repository and package; experimental outcomes depend on the selected recipe. This entry describes the source audit of 14 September 2026; historical measurements retain their original dates.
Mechanism
Keep one live model with active adapter and accumulated residual state. Inference and queued training operate under an ordering/mode-lock contract. Training returns a commit token; a later inference request can require visibility of that commit. Objectives are independent callables composed by the trainer.
Implementation and controls
The daemon lives at ~/ht/trainfer; the research/teaching client lives at ~/ht/cont. The objective registry, state module, trainer, and snapshot implementation define the actual available methods. Trainfer learning methods catalogues every registry entry, the current client mechanisms, the recovered historical arms, and concrete proposals.
Evidence and evaluation
The project has primitive tests, ranking benchmarks, local learning/retention experiments, and several negative campaign findings. A functioning service is established separately from the research goal of consistent, sample-efficient continual learning.
Limitations and interpretation
A single shared model is not unrestricted concurrent GPU execution: ordering and mutual exclusion are part of correctness. Status tables and plans may describe earlier repository names or intended methods; use source-pinned articles for exact implementation boundaries.
Sources
- trainfer: README.md — checkout audited
1c6391f3773b. - trainfer: trainfer/objectives/__init__.py — checkout audited
1c6391f3773b. - trainfer: trainfer/engine/train.py — checkout audited
1c6391f3773b. - trainfer: trainfer/state.py — checkout audited
1c6391f3773b. - cont: AGENTS.md — checkout audited
87946914c7b9.