tars / NOTICE
aneuwirth's picture
Tars 0.5
037fdd3 verified
Raw History Blame Contribute Delete
1.98 kB
Tars
Copyright 2026 The Tars Authors
https://github.com/andrewneuwirth/Tars
This product is based on Von:
Von, the open-source System One decision model
Copyright 2026 Victor Hugo Panisa (Victor Hugo)
https://github.com/wfzyx/von
Licensed under the Apache License, Version 2.0
Von's model uses the ModernBERT encoder:
ModernBERT, Answer.AI and LightOn (https://github.com/AnswerDotAI/ModernBERT)
Licensed under the Apache License, Version 2.0
What Tars uses from Von
- Von's model class (von.models.option_marker.OptionMarkerModel), its training
collator and loss (training/train_option_marker.py) and its synthetic
general-reasoning generator (training/generate_synthetic_decisions.py).
These are imported from a Von checkout at run time, not copied.
- Von 1.2's weights are the starting point of every Tars fine-tune.
- pipeline/train.py and tars/model.py adapt parts of Von's training loop and
model loading; each file's header says which parts.
What Tars changed or added
- Three decision questions for commands, built from a skill catalog:
split (does a new step start here?), route (which family?) and skill
(which skill in that family?) (tars/questions.py).
- A planner that chains the three questions into steps (tars/planner.py)
and an HTTP server (tars/server.py).
- A full fine-tuning pipeline: label generation, dataset build with speech
recognition noise and general-reasoning replay, time-budgeted training on
Apple GPUs or CUDA, and a promote/keep evaluation (pipeline/).
- A macOS app for defining a catalog, collecting labels from your own AI
provider or any MCP client, reviewing them, training and comparing models (app/).
Training labels
Tars's own training phrasings were written by an LLM (Anthropic's Claude,
through the Claude Code CLI). Labels you make with the app come from
whatever provider you choose. Check your provider's terms before
redistributing generated datasets.