Hi ABC team,
I built TraceOS (aaa-mvc/traceos), a lightweight experiment runtime that wraps ABC's training and evaluation pipeline.
What TraceOS does:
experiment run bottle.yaml → one command: train + evaluate + report + analyze
Standardized output structure per run (report.html, analysis.json, events.jsonl)
Run registry for experiment history
Cross-run comparison
Capability scoring + failure taxonomy + recommendations
How it relates to ABC:
The ABC Adapter wraps train.py and eval_policy.py via subprocess
Zero lines of ABC source code are modified (verified by git diff)
ABC's DiT model, DINOv3 backbone, CLIP encoder, and MuJoCo sim are used as-is
The paper and dataset are prominently acknowledged in our README
Status:
v0.1.0 released under Apache 2.0
97 tests passing
Mock adapter available for instant demo (no GPU needed)
Thank you for open-sourcing ABC-130K — it made this work possible. Any feedback is welcome.
Hi ABC team,
I built TraceOS (aaa-mvc/traceos), a lightweight experiment runtime that wraps ABC's training and evaluation pipeline.
What TraceOS does:
experiment run bottle.yaml → one command: train + evaluate + report + analyze
Standardized output structure per run (report.html, analysis.json, events.jsonl)
Run registry for experiment history
Cross-run comparison
Capability scoring + failure taxonomy + recommendations
How it relates to ABC:
The ABC Adapter wraps train.py and eval_policy.py via subprocess
Zero lines of ABC source code are modified (verified by git diff)
ABC's DiT model, DINOv3 backbone, CLIP encoder, and MuJoCo sim are used as-is
The paper and dataset are prominently acknowledged in our README
Status:
v0.1.0 released under Apache 2.0
97 tests passing
Mock adapter available for instant demo (no GPU needed)
Thank you for open-sourcing ABC-130K — it made this work possible. Any feedback is welcome.