FluidInference/laya-english-coreml

Laya English root checkpoint for Core ML

This repository hosts the English root checkpoint of convaiinnovations/laya, pinned at 1c5edc17a7acd8701df6fc341c0d179f1c62c982. It is a 421,293,827-parameter ModernBERT-large model with trained choice, score, noul and action heads. This is separate from FluidInference/laya-coreml, which hosts the roughly 322M multilingual checkpoint.

The L128 and L512 packages are fixed-shape FP16 Core ML programs with 32 option slots, targeting iOS 17/macOS 14 or newer. Use L128 when the full rendered input fits 128 tokens and L512 for 129–512 tokens. Inputs: input_ids [1,L] int32, attention_mask [1,L] int32, marker_map [1,32,L] float32, question_type [1,3] float32. Outputs: raw logits, uncalibrated probabilities, and action_probabilities. The host must use rl_agent_config.json to apply the English model's type- and option-count-specific temperatures to logits. The included standalone calibration.py contains the temperature and softmax helper; preprocessing.py needs the upstream laya Python package for rendering. Requests that exceed 32 options or the installed sequence bucket must be rejected or explicitly handled under the upstream API's truncation policy.

The original Laya root config has max_len=512 and head_max_len=192. Native PyTorch parity for both packages is recorded in verification-english-L128.json and verification-english-L512.json. On Apple M5 Pro, the validated automatic/GPU route matched all 16 fixture choices at both lengths. Worst calibrated probability errors were 0.00225 (L128) and 0.00425 (L512); median model call times were 7.79 and 20.79 ms. Forced CPU+ANE also matched every choice, but its worst calibrated probability error was 0.0355, failing the predeclared 0.02 gate on a 20-option case. The English checkpoint uses a temperature of 0.10058 there, which amplifies small logit differences. Use .all for these packages; CPU+ANE is not validated for calibrated outputs. The source lock and conversion reports record SHA-256 hashes and parameter count.

The Decision Index tracker reports Laya at 16.39, but its historical adapter's selected checkpoint has not yet been authenticated. This English artifact must not be presented as a verified reproduction of that score. See FluidInference/mobius for conversion source and upstream Laya for the Apache-2.0 model and its original evaluation.