SemiKong: Curating, Training, and Evaluating A Semiconductor Industry-Specific Large Language Model
Paper • 2411.13802 • Published • 2
Datasets:
session stringclasses 10
values | lot stringclasses 10
values | program stringclasses 1
value | yield_gap_pct float64 5.15 8.3 | parametric_escape_count float64 8 17 |
|---|---|---|---|---|
FT-AU5-260218-17 | AU5-260218-17 | FT-PROD-R8 | 5.15 | 8 |
FT-CA7-260218-24 | CA7-260218-24 | FT-PROD-R8 | 5.5 | 9 |
FT-DE12-260218-09 | DE12-260218-09 | FT-PROD-R8 | 5.85 | 10 |
FT-EM3-260218-31 | EM3-260218-31 | FT-PROD-R8 | 6.2 | 11 |
FT-FP28-260218-11 | FP28-260218-11 | FT-PROD-R8 | 6.55 | 12 |
FT-IO14-260218-06 | IO14-260218-06 | FT-PROD-R8 | 6.9 | 13 |
FT-LA40-260218-22 | LA40-260218-22 | FT-PROD-R8 | 7.25 | 14 |
FT-ME65-260218-14 | ME65-260218-14 | FT-PROD-R8 | 7.6 | 15 |
FT-PO22-260218-19 | PO22-260218-19 | FT-PROD-R8 | 7.95 | 16 |
FT-SU16-260218-27 | SU16-260218-27 | FT-PROD-R8 | 8.3 | 17 |
SemiKongBench-100
One hundred independently authored synthetic semiconductor operations tasks: ten workflows across ten frozen fabs. Each task includes 20 agent-visible assets, a task-local SQLite world, 38 provider-shaped tools, an oracle trajectory, before/after snapshots, and a deterministic 100-point verifier.
The model leaderboard is intentionally empty until a complete version-pinned 100-task model run exists. Release qualification executes oracle, replay, and six negative-control policies; controls are not model results.
Provenance and relationship to SemiKong
https://github.com/aitomatic/semikong and the SemiKong paper are public design anchors. This benchmark is not produced by, endorsed by, or an official scorecard of Aitomatic, SemiKong, or the AI Alliance. No SemiKong training/evaluation corpus or model weights are redistributed. Every task, value, file, tool, answer, and trajectory here is new synthetic content from Blobfish AI.
Dataset content is CC BY 4.0. Executable code is Apache-2.0 in the source repository.
- Downloads last month
- -