DanielJeongsooLee/seal-5comp-d9-20260923

SEAL 5-component D9 recollected dataset

A fixed, group-disjoint Train/Validation/Test release of 133,263 directly simulated circuit-duty responses, 14,807 typed circuits, and 3,060 connection-structure groups. Every circuit has five components selected from Type-1/Type-2 switches (Sa/Sb), inductors (L), and capacitors (C), with external terminals Vin, Vout, and GND.

The circuit topology source is LaMAGIC / LaMAGIC-dataset. These are our preserved common group assignments and new NGSpice responses under one pinned simulation profile, not the upstream training/validation split labels. split in the public Parquet files is the authoritative common split. The _source.partition and old split fields in raw archives describe historical source provenance only.

Split	Connection groups	Typed circuits	D9 response rows
train	2,000	9,528	85,752
validation	360	1,820	16,380
test	700	3,459	31,131

The full local allocation has 3,803 groups; its 743-group Reserve partition remains unused and is not part of the published response tables. All typed variants belonging to a common connection group remain in the same partition. Switch types are preserved; the connection group identifier is not an exact typed-circuit identity.

Loading

from datasets import load_dataset

responses = load_dataset("DanielJeongsooLee/seal-5comp-d9-20260923", "responses")
identities = load_dataset("DanielJeongsooLee/seal-5comp-d9-20260923", "identities")
teachers = load_dataset("DanielJeongsooLee/seal-5comp-d9-20260923", "teachers")
contexts = load_dataset("DanielJeongsooLee/seal-5comp-d9-20260923", "e1_k500_contexts")

import json
topology = json.loads(responses["train"][0]["topology_json"])

The default configuration is responses. teachers contains the original D9-supervised duty teachers for Train and Validation only, including an explicit eligible flag. The Task models in the accompanying E1 K500 release use the fixed filtered context subset, not every full-data teacher row.

Fields and graph representation

• topology_hash: preserved exact typed-circuit identifier.
• group_hash: preserved connection group used to prevent split leakage.
• topology_json: lossless JSON topology with components, net_by_port, and terminals. Component entries contain name, kind, and two ports.
• duty: stored duty, unchanged by recollection or publication.
• gamma: measured output/input voltage ratio; efficiency: measured efficiency.
• status: simulation outcome. eligible_physical_supervision identifies valid finite supervision rows. Invalid responses are retained and are not silently relabeled as physical zero targets.
• utility_targets, utilities: ten matched target/value entries.
• simulation_profile_hash, simulation_result_sha256: physical-condition and result identity.

Each circuit has exactly nine distinct stored duties: anchors 0.1, 0.3, 0.5, 0.7, 0.9, plus one preserved duty in each adjacent open interval. Interior duties are not resampled. The original raw/<split>/*.jsonl.gz files decompress to byte-identical release JSONL files. They retain original metadata, historical legacy_response fields, and local provenance paths; the public response columns use the current recollected response only. Legacy values are not mixed into the current labels. Detailed simulator result files and waveforms referenced by those local paths are not included in this compact tabular release.

Simulator and scoring

The pinned profile is lamagic_official_benchmark_v1 / scratch-lamagic-benchmark-v1, with the original official simulator source and NGSpice binary identities recorded in the model bundle's simulator code and provenance. Physical conditions:

Condition	Value
Input voltage / switching frequency	100 V / 1 MHz
Inductance	100 microhenry
Capacitors / output capacitor	10 microfarad
Input resistance / load resistance	0.1 ohm / 50 ohm
Transient	tran 10n 4000u
Timeout	500 seconds per call
Measurement	Preserved official waveform extraction and validity rules, final 50-cycle window

Targets are [-3, -2, -1.5, -1, -0.5, 0.25, 0.5, 1.5, 2, 3]. For valid finite responses:

[ U = \max(\eta,0),1.1^{-\left(15\frac{\gamma-\gamma^*}{|\gamma^*|}\right)^2}. ]

Completed invalid responses, timeouts, and simulator errors have Utility zero. Their status remains explicit. They are not valid physical training labels.

E1 K500 subset and model selection

e1_k500_contexts contains 2,556 Train, 1,643 Validation, and 2,935 Test circuit-target contexts. Eligibility requires at least five valid finite D9 responses and a valid-response Utility span of at least 0.02. Train additionally keeps the top 500 eligible contexts per target by Utility span, including all ties at the cutoff. There is no top-500 truncation on Validation or Test. The fixed Test subset covers 2,248 typed circuits and 483 groups; it is narrower than the full 700-group Test split in this dataset.

The public context tables contain identities and targets only. They omit model-proposed Test duties. Original Train/Validation cohort files, including duty supervision, are preserved under raw/e1_k500_contexts/.

The paired model release is SEAL E1 validation-best models. Physical models are selected using Validation physical-response MSE; Task models are selected using direct-NGSpice Validation Utility. Test was not used for checkpoint selection. Training uses a strict Train/Validation-only manifest; its unchanged original SHA-256 is recorded in dataset_catalog.json.

Provenance and validation

dataset_catalog.json maps portable Parquet files to original source records and hashes. provenance/original/ contains unchanged historical audit/manifests; absolute paths there identify the original research environment and are not portable download destinations. Public copies do not replace original provenance or alter the active E1 run.

Publication checks verified original file hashes, all split/group counts, disjoint group and typed-circuit sets, preserved nine-duty support, simulation profile identity, current Utility values, byte-identical decompression of original JSONL, and Parquet round trips. The model release records the same training-data manifest hash.

This release enables model training, duty prediction, and fixed-split evaluation. It does not establish generalization to unseen connection structures from changing an MCTS seed, and the initial release contains no final E1 Test performance claim.

Attribution and reuse

Please cite the original LaMAGIC work when using its circuit topology source, and cite this release identifier/revision for the reconstructed split and common-condition recollection. The upstream sources retain their attribution and terms. No new license is assigned by this packaging step.