Update spellchecker model card and license

README.md

@@ -48,8 +48,7 @@ base_model:
 
 # LFM2.5-Encoder-350-Spellchecker
 
-A full fine-tune of [LFM2.5-Encoder-350M](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M) with a subword-level **GECToR-style grammatical-error-correction tagger**. 
-It covers grammar, spelling, punctuation, and casing in English. 
+A full fine-tune of [LFM2.5-Encoder-350M](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M) with a subword-level **GECToR-style grammatical-error-correction tagger** (~357M params). It predicts a per-BPE-piece edit tag and applies edits iteratively (`KEEP / DELETE / REPLACE / APPEND`), covering grammar, spelling, punctuation, and casing across 11 languages. Selected automatically as the best checkpoint by the **MASTER composite** (a weighted aggregate over native, learner, robustness, precision and multilingual evals; leaderboard rank **#1**, source run `en/stage3_ptv2`).
 
 Find more details about our encoders in our [blog post](https://www.liquid.ai/blog/lfm2-5-encoders).
 

config.json

@@ -7,7 +7,7 @@
     "AutoConfig": "modeling_gectagger.GecTaggerConfig",
     "AutoModel": "modeling_gectagger.GecTaggerForGEC"
   },
-  "encoder_name": "LiquidAI/LFM2.5-Encoder-350M",
+  "encoder_name": "LiquidAI/mlm_phase2_bidir2_step140800",
   "num_tags": 128802,
   "hidden_size": 1024,
   "tie_replace": true,
@@ -18,4 +18,4 @@
   "qat_group_size": 32,
   "dropout": 0.1,
   "torch_dtype": "float16"
-}
+}

reranker/README.md

@@ -0,0 +1,28 @@
+# Edit reranker (EditScorer)
+
+Per-edit filter that lets the tagger over-generate (negative `keep_confidence`) and restores
+precision by scoring each proposed edit — Tier 1b of the architecture roadmap. Pairwise scorer:
+encodes the EDITED sentence and the ORIGINAL sentence (same LFM2.5 backbone as the tagger, encoder
+seeded from the shipped checkpoint) and pools both edit spans.
+
+**Shipped operating point: `keep_confidence=0.0`, `tau=0.3` -> MASTER composite 0.6587**
+(tagger alone at fixed thresholds: 0.631). fp_clean (clean-text false-flag guard) = 1.0.
+Pooled English M2 at this point: P=0.6636 R=0.3076 (tagger baseline: P=0.544 R=0.269 — the
+reranker buys ~+8pts precision at equal recall).
+
+## Usage (with the `spellchecker` repo)
+
+```python
+from spellchecker.predict import GecPredictor
+from spellchecker.reranker import EditScorer, RerankedPredictor
+
+base = GecPredictor.from_pretrained("<tagger>", device="cuda",
+                                    keep_confidence=0.0, max_iter=4)
+scorer = EditScorer.load("<this folder>").to("cuda")
+pred = RerankedPredictor(base, scorer, tau=0.3)
+pred.correct(["She go to school every day ."])
+```
+
+Training: candidates harvested from the tagger's own over-generations on real learner data
+(bea/fce/lang8) + clean anchors, labelled by gold-edit match. See eval report in
+`operating_point.json` and `docs/ARCHITECTURE_ROADMAP.md` (Phase 2) in the training repo.

reranker/scorer_config.json

@@ -1,6 +1,6 @@
 {
-  "encoder_name": "LiquidAI/LFM2.5-Encoder-350M",
+  "encoder_name": "LiquidAI/mlm_phase2_bidir2_step140800",
   "hidden_size": 1024,
   "mode": "cross",
   "type_feature": true
-}
+}