slic-agent/slic-data

SLIC data

Analog circuit optimization tasks for multi-turn LLM agents (65 nm bulk CMOS, open predictive model card, ngspice), and the transistor model cards for the technology-node experiments. The code that uses them is the slic-agent repository; slic data downloads the folders below into the layout it expects.

folder	content	slic data
standard/	training pool, IID and OOD test sets (and the IID subset evaluated during training) used by the paper's main runs	--family standard
hard/	harder tasks on the same topologies: one specification starts far from its target (not used in the paper)	--family hard
nodes/	the standard IID subset / OOD test tasks rebuilt for 22, 32, 45, 90 and 130 nm	--family nodes
ptm-cards/	predictive technology model cards (Spectre syntax) for those nodes	--cards
topologies/	one folder per topology: schematic.svg / .png and netlist.scs (the parameterised circuit every task row of that topology carries)	—

topologies.csv lists the 41 topologies (34 in-distribution, 7 held out for the OOD tests): circuit kind, signal mode, device count, number of tunable parameters, and links to its schematic and netlist.

Each task folder has its own README (files, row counts, columns), a stats.json (per file and per topology: row count, valid / invalid starts, and mean / median / min / max of the initial metrics and of the target specifications) and SHA256SUMS; the top-level SHA256SUMS covers every file in the repository.

The dataset viewer above shows the standard, hard and nodes task sets (one subset per family, one split per file); every row is one task, with the circuit, the action space, the goal and the simulated starting point in extra_info.

Version. The paper's results were computed on an earlier version of these task sets that contained errors in some circuits and in the measurement; the sets here are the corrected ones, so numbers on them are not directly comparable to the paper's tables.

Licenses. The task sets (standard/, hard/, nodes/) are released under CC BY 4.0. The model cards in ptm-cards/ are the Predictive Technology Model cards of Arizona State University (https://ptm.asu.edu), converted to Spectre syntax with unchanged parameters, and are redistributed under the terms of their origin (free for research use; see the PTM site). See LICENSE.md.