LFM2.5-Encoder-350M-Spellchecker — LiteRT

LiquidAI/LFM2.5-Encoder-350M-Spellchecker converted to LiteRT (.tflite) for on-device inference. A GECToR-style two-head tagger that corrects misspellings and grammar token by token, fully offline (demo Space).

File	Recipe	Size
LFM2.5-Encoder-350M-Spellchecker_wi8fc.tflite	int8 dynamic-range (linears + embedding + tied vocab heads, convs float)	429 MB	mobile + desktop (iPhone-verified bit-exact, 64 ms)
LFM2.5-Encoder-350M-Spellchecker_fp16.tflite	fp16 weights, float compute	847 MB	desktop — phone memory limits (XNNPACK per-signature fp32 unpacking)

Signature

gec_128 (S = 128, batch 1, right-padded; the base model's own decode uses max_len 128): input_ids int32 [1, 128] (prepend the tokenizer BOS as the sentence anchor), attention_mask int32 [1, 128] → two outputs, both zeroed at padded positions:

Output	Shape	Meaning
label_logits	float32 [1, 128, 128802]	per-token edit tag: 0 $KEEP, 1 $DELETE, 2..2+V $REPLACE_<piece>, 2+V.. $APPEND_<piece> (V = 64400 BPE pieces)
detect_logits	float32 [1, 128, 2]	P(token is part of an error) gate

Host-side decode is the base repo's algorithm: argmax the tags, gate by softmax(detect)[1] >= min_error_prob, apply the edits, and iterate (≤3 passes) until the text stops changing. The base repo also bundles an optional PyTorch reranker for its published maximum-precision operating point — that stays host-side/desktop; this artifact covers the tagger (a fully supported mode of the base model's .correct()).

Verification

Task-level parity vs the PyTorch reference ("She go to school every day ." → single $REPLACE on "go"): fp32, fp16 and int8 all produce the identical edit (same position, same replacement piece, detect head agreeing). On an iPhone 17 Pro the int8 file reproduces the desktop outputs bit-exactly on both heads — including the full [1, 128, 128802] label tensor — (cosine 1.000000, max diff 0.0) at 64 ms per pass (6 threads, XNNPACK).

License

LFM Open License v1.0 (see LICENSE, unchanged from the base model). Note the license's commercial-use threshold (Section 5). This repository redistributes converted Derivative Works of LiquidAI/LFM2.5-Encoder-350M-Spellchecker with modification notices per Section 4; all credit for the model to Liquid AI.