ChipLaya v1.0.1

ChipLaya v1.0.1, a code-only release. The model and its weights are unchanged: they are the v1.0.0 assets (typed-decisions.safetensors 08155a26…, typed-decisions.pt f5df4faa…), which ChipLaya.pretrained() downloads and verifies.

• weights.fetch removes partial or wrong downloads on any failure, and cached files get normal (umask) permissions.
• A package release may reuse earlier weights; chiplaya.weights.RELEASES lists weights releases.
• scripts/check_chipjev.py checks ChipJev's pin strictly: the weights must be the pinned weights release, and the tag must match the version.
• Model card corrections. A free-text request takes two passes. The CLI's reported timing is explained. The card also says what is verified when, adds offline=False to the Python example, notes the discarded development candidates, and fixes the label builder's location.

pip install "chiplaya[runtime] @ git+https://github.com/lab-emi/ChipLaya@v1.0.1"

ChipJev (lab-emi/ChipJev) vendors this release. ChipLaya is an independent fine-tune by the EMI Lab, not affiliated with or endorsed by Convai Innovations.

ChipLaya v1.0.0

ChipLaya v1.0.0 — the decision model of ChipJev (live demo), released on its own.

ChipLaya is the Laya multilingual decision model by Convai Innovations, run through a PyTorch port of Laya-MLX, with 35M of its 322M parameters fine-tuned on circuit-design development data. It answers the typed questions of an amplifier design request (class, objective, input stage, device, stage count, later stages, compensation, buffer) in one batched pass of about 8 ms on an RTX 4090.

These are exactly the weights behind ChipJev's frozen PTM 45 nm and SKY130 studies (experiments/ptm45/typed-decisions.pt).

Asset	SHA-256
typed-decisions.safetensors (56 FP16 tensors, 34,831,873 parameters)	08155a2669684494318da6b309e2c06e97e341dc6abce16ddd2e546306471b71
typed-decisions.pt (the same tensors; the archival file ChipJev pins)	f5df4faaeb1e2721657e5200293f7667ca525b8f86c1c8d3c3ab14193baeb8d5
LICENSE, NOTICE, THIRD_PARTY_NOTICES.md, Apache-2.0.txt, MIT-mmBERT.txt	see SHA256SUMS

Base checkpoint (downloaded separately, pinned): aac6fef/laya-multilingual-mlx @ f2b4faf5 (model.safetensors SHA-256 7fc5834a…), converted from convaiinnovations/laya-multilingual @ 052592a1; encoder jhu-clsp/mmBERT-base (MIT).

pip install "chiplaya[runtime] @ git+https://github.com/lab-emi/ChipLaya@v1.0.0"
chiplaya "Design a two-stage op-amp with the highest gain-bandwidth product" --vdd 1.2 --load 100

Evaluation, training data and limitations: see the model card. ChipLaya is an independent fine-tune by the EMI Lab (TU Delft), not affiliated with or endorsed by Convai Innovations, and unrelated to TypeSafe AI's Jev. Code and weights: Apache-2.0; attribution in NOTICE.