The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
ATC tree20 runs
Training artifacts for the Abstract Token Curriculum (ATC) twin-random-tree (size-20) experiment. Task: two isomorphic random trees of 20 nodes each (~38 edges), root in one of them, two candidate leaves shown; the model must name the reachable one. Hop distance d = 4..15. Answer-only supervision -- no intermediate state is ever labelled or generated. Chance = 0.5.
Model: 2-layer / 8-head / 768-dim GPT-2 trained from scratch. Recipe: CE loss, curriculum=ring, candidate dropout 0.75, lr 3e-5, max_latent_stage=16, per-instance budget=depth-1, uniform_prob=0.5.
Two findings
1. Training-set size is what clears stage 0. Identical arms at n_train=10000 never left stage 0 in ~247 epochs (held-out stage-0 accuracy pinned at ~0.49) while their training loss fell to 0.64 -- memorisation. At n_train=100000 every arm cleared the stage-0 gate within ~25 epochs and the task was solved to full depth (best arm 1.000, stage 16).
2. Truncated backprop needs the retentive fallback. backprop_last x staging is an interaction, not two main effects (sustained accuracy, mean of last 10 evals):
| backprop full | backprop_last=5 | |
|---|---|---|
| threshold (monotone) | 0.999 / 0.998 | 0.503 / 0.506 |
| revisit (fallback) | 0.996 / 0.993 | 0.950 / 0.899 |
With full gradients the staging rule is irrelevant. With truncated gradients, monotone collapsed from 0.98 to chance at the stage 7 to 8 transition (train loss 1.09 to 2.95) and never recovered, while revisit fell back, repaired, and held ~0.95. Reading: full backprop maintains early stages implicitly; truncation removes that path, so only explicit repair can fix the decay.
Layout
d100k folders -- the 8 arms at n_train=100000: checkpoint_latest.pt (full resumable state), epoch snapshots, config.json, metrics.jsonl. control-10k folders -- the 8 arms at n_train=10000, metrics and config only (their weights sit at chance and are not worth distributing). These are the data-starvation control.
Single seed per cell. See docs/FINDINGS.md section 13 in the source repo.
- Downloads last month
- 78