parrotflow-syrinx

Core ML models for ParrotFlow, built on jhu-clsp/mmBERT-small.

parrotflow-syrinx-fp16.mlpackage

One fp16 ML program (macOS 15) with three functions. They share one copy of the base weights: the 256,000 x 384 table and blocks 0-21 are stored once, in fp16. The package is 283,340,586 bytes.

Function	Inputs	Outputs	What it is
`mask`	`input_ids`, `attention_mask`: int32 (1, 64)	`logits`: fp16 (1, 64, 256000)	The mmBERT-small masked-LM program from `znaat/mmbert-small-coreml` at `a983542`, added unchanged. Bit-equal to it on 641 of 641 probes.
`parse`	`input_ids`, `attention_mask`: int32 (1, 256); `g_qkv_A`, `g_o_A`, `s_qkv_A`, `s_o_A`: fp16 (12, 4, 384); `g_qkv_B`: fp16 (12, 1152, 4); `g_o_B`: fp16 (12, 384, 4); `s_qkv_B`: fp16 (12, 1152, 4); `s_o_B`: fp16 (12, 384, 4)	`g_hidden`, `s_hidden`, `s_hidden10`: fp32	Blocks 0-11 with two rank-4 LoRA adapters (G and S) on `attn.Wqkv` and `attn.Wo`, kept unmerged: `x @ W.T + scaling * (x @ A.T) @ B.T`, scaling 2.0. The adapters are inputs, so a new fine-tune needs no re-export.
`ner`	`input_ids`, `attention_mask`: int32 (1, 256)	`n_hidden`: fp32	Blocks 0-10 with a rank-8 LoRA adapter (constants), the raw output of block 10.

The default function is parse. Set MLModelConfiguration.functionName to pick another one. Run it on .cpuAndGPU.

The tokenizer is mmBERT-small's: tokenizer.json, tokenizer_config.json and config.json in this repository.

Sources:

Base: jhu-clsp/mmBERT-small, snapshot abc32620dd4f6ab06f5fbe905dc25f310618e09f.
G adapter: the grammar adapter of the Swift disfluency v1 weights.
S adapter: the syntax adapter, t-silver seed 42.
NER adapter: the 13-type NER run F-both seed 43, epoch 8.

Tools: Python 3.13, torch 2.7.0, coremltools 9.0, transformers 4.48.3.