README.md

metadata

license: mit
tags:
  - atc
  - coconut
  - continuous-cot
  - graph-reasoning
  - checkpoints

ATC tree10 runs

Training artifacts for the Abstract Token Curriculum (ATC) twin-random-tree (size-10) experiment: a 2x2x2x2 factorial over loss (CE vs sum-of-log), reset_optimizer (on/off), staging (threshold/monotone vs revisit), and uniform_prob (0.1/0.5), plus related arms.

Each subfolder is one run and contains:

• checkpoint_latest.pt - full resumable state (weights + optimizer + curriculum position)
• checkpoint_<n>.pt - epoch-stamped snapshot(s)
• config.json - the run configuration
• metrics.jsonl - per-epoch metrics (eval/acc, stage, per-stage acc, losses)

Model: 2-layer / 8-head / 768-dim GPT-2 trained from scratch on symbolic reachability graphs.