snow-leopard/opal-results

OPAL run records and checkpoints

OPAL (OPtimizer with an Agent in the Loop) is a PyTorch optimizer wrapper that lets a coding agent read training observations and rewrite optimizer and analysis programs during training. This dataset contains the selected release run records and retained model weights.

Code and reproduction instructions · Project and blog

Each setting is a single run. These records do not establish variation across seeds. Static runs use the fixed recipe; OPAL runs allow agent changes. The two runs under codex_backend_experimental/ use the experimental Codex backend with gpt-6.1-sol at high effort.

Lower is better for FID, validation loss and FLORES-200 bits per sentence; higher is better for accuracy. Track 3 validation loss is in nats per token. Values below are rounded for display; each result.json retains the recorded precision and timing definitions.

Example	Run	Role	Metric	Value
raev2	dopt-raev2-b20-opal-499fddb-rs2	OPAL	FID-50K (EMA)	3.19
raev2	dopt-raev2-b20-official-v1	static	FID-50K (EMA)	3.39
goldfish amh	dopt-gf-amh-opal-499fddb-official-a2	OPAL	FLORES-200 bits/sentence	81.57
goldfish amh	dopt-gf-amh-official-v1	static	FLORES-200 bits/sentence	83.48
goldfish swa	dopt-gf-swa-opal-499fddb-official-a2	OPAL	FLORES-200 bits/sentence	77.25
goldfish swa	dopt-gf-swa-official-v1	static	FLORES-200 bits/sentence	79.37
goldfish kir	dopt-gf-kir-opal-499fddb-official-a2	OPAL	FLORES-200 bits/sentence	69.03
goldfish kir	dopt-gf-kir-official-v1	static	FLORES-200 bits/sentence	70.81
Track 3 zho hans	t3-o02d-3k-499-rs-p0	OPAL	validation loss raw	4.21400
Track 3 zho hans	s3000-f3d-r02	static	validation loss raw	4.25703
Track 3 zho hans	t3-o36d-3k-447-p0	OPAL	validation loss raw	4.12857
Track 3 zho hans	s3000-f3d-r36	static	validation loss raw	4.18447
Track 3 tha thai	t3-o02f-3k-499-rs-p0	OPAL	validation loss raw	4.72078
Track 3 tha thai	s3000-f3f-r02	static	validation loss raw	4.77445
Track 3 tha thai	t3-o36f-3k-499-p0	OPAL	validation loss raw	4.66083
Track 3 tha thai	s3000-f3f-r36	static	validation loss raw	4.71995
Track 3 kir cyrl	t3-o02c-3k-499-rs-p0	OPAL	validation loss raw	3.94820
Track 3 kir cyrl	s3000-f3c-r02	static	validation loss raw	4.00449
Track 3 kir cyrl	t3-o36c-3k-499-p0	OPAL	validation loss raw	3.85329
Track 3 kir cyrl	s3000-f3c-r36	static	validation loss raw	3.91228
Track 3 fineweb quarter	t3-og-499-rs2-p0	OPAL	validation loss raw	3.30782
Track 3 fineweb quarter	f2e-r36-lr1-loss	static	validation loss raw	3.31039
overfit label10	op-ofit-499-b	OPAL	test accuracy (%)	80.00
overfit label10	op-ofit-499-te2	OPAL	test accuracy (%)	80.42
overfit label10	ssln-s-label10-3e3	static	test accuracy (%)	73.53
cifar10 resnet	op-cifar-499-b	OPAL	test accuracy	0.9456
cifar10 resnet	op-cifar-499-plain	static	test accuracy	0.9449
gpt small	op-gpt-499-b	OPAL	validation loss	4.3419
gpt small	op-gpt-499-plain	static	validation loss	4.4763
diffusion mnist	op-diff-499-b	OPAL	validation loss ema	0.1623
diffusion mnist	op-diff-499-plain	static	ema validation loss	0.1696
mnist mlp	mnist-plain	static	test accuracy	0.9850
mnist mlp	mnist-opal	OPAL	test accuracy	0.9863
gpt small	op-gpt-sol-b	OPAL	validation loss	4.3125
diffusion mnist	op-diff-sol-b	OPAL	validation loss ema	0.1687

Track 3 OPAL runs also record Tail-EMA validation losses in secondary_metrics; the table uses raw-weight validation loss for the static/OPAL comparison. The quarter-batch static f2e-r36-lr1-loss ends at 3.31039 in its own retained stdout.

Start with result.json in each run folder. It describes the metric, baseline, command, software revision, hardware, timing scope, agent decisions/cost when recorded, missing artifacts and weight hashes. artifacts/ holds the original per-run metrics, logs, decision observations and messages, generated programs and reports when retained. provenance/ holds relevant saved code/configuration. Available contents vary by run. Each run has SHA256SUMS; the root MANIFEST.json inventories the published files.

Checkpoints contain tensor state dictionaries only. Model and EMA states are retained when present; optimizer, agent and other non-tensor metadata are omitted. Original and published SHA256 values are recorded in each run’s result.json. RAEv2-B weights are from the final epoch-20 checkpoint at step 25020; the reported FID used EMA at step 25000, whose exact snapshot was not separately retained. Goldfish includes the final checkpoint for each listed run. Overfit best.pt is selected by development accuracy: the static checkpoint is from update 40771; the two OPAL checkpoints are from the final update 42120.

Track 3 has records only because no model or Tail-EMA checkpoints were retained. The ordinary small scripts and MNIST MLP likewise have no retained weights. The overfit and two experimental Codex runs include the weights that were saved. Training data are not included.

This release contains selected finished runs. Internal aggregate research reports, deployment orchestration and disposable caches are outside the package. Paths, storage locations, hostnames, infrastructure identifiers and personal identifiers were replaced with neutral placeholders in copies; source bundles were preserved privately. See REDACTION.md for the procedure and REDACTION.json for the audit.

The run records and exported checkpoints in this repository are released under the MIT license. See LICENSE.