Add v2.5 no-aug ablation checkpoint

Preserves the first training run (augmentation off) for ablation reference. Superseded by augmented production checkpoint.

README.md

@@ -0,0 +1,18 @@
+---
+license: mit
+tags: [ielts, automated-essay-scoring, deberta-v3, evalguide, ablation]
+---
+
+# EvalGuide IELTS AES v2.5 — no-aug ablation (superseded)
+
+This commit preserves the first v2.5 training run where synonym augmentation was **disabled** (NLTK WordNet not loaded).
+
+**Superseded by the augmented + calibrated production checkpoint in the next commit.**
+
+| Field | Value |
+|-------|-------|
+| Variant | No-aug ablation |
+| Gold harness raw QWK | 0.8595 |
+| Augmentation | Disabled (unintentionally) |
+
+See repo HEAD for the production model card.

ielts_v2.5_base_en_10ep.weights.h5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b224be1ff02811af7c7b67d8ab750a0c43506f3060f3b4a081060bd86e85c9e
+size 3696208736

ielts_v2.5_base_en_10ep_calibration.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f28235d93b1e494db459f98f1f4f194669ef0ad4d5b4afb9c4a339a0b445cc8c
+size 1750

ielts_v2.5_base_en_10ep_config.json

@@ -0,0 +1,54 @@
+{
+  "version": 2.5,
+  "preset": "deberta_v3_base_en",
+  "sequence_length": 512,
+  "num_band_scores": 19,
+  "criteria": [
+    "Task Response",
+    "Coherence and Cohesion",
+    "Lexical Resource",
+    "Grammatical Range and Accuracy"
+  ],
+  "training_params": {
+    "batch_size": 8,
+    "epochs": 10,
+    "learning_rate": 2e-05,
+    "label_smoothing": 0.1,
+    "augment_ratio": 0.1,
+    "synthetic_weight": 0.15,
+    "synthetic_csv": "/home/youwen/AEGIS/evalguide_client/backend/data/ielts_synthetic_cleaned.csv",
+    "initial_variance_target": 2.0,
+    "final_variance_target": 2.7,
+    "calibration_enabled": true,
+    "bias_correction_enabled": true
+  },
+  "bias_correction_offsets": [
+    -4.853433987432378e-17,
+    1.553098875978361e-16,
+    1.9413735949729513e-17,
+    8.736181177378281e-17
+  ],
+  "final_metrics": {
+    "train_loss": 0.26780974864959717,
+    "val_loss": 0.41971996426582336,
+    "binary_acc": 0.12838737666606903,
+    "mae": 2.637578248977661,
+    "val_binary_acc": 0.0,
+    "val_mae": 2.3622758388519287,
+    "val_qwk": 0.5819463133811951,
+    "val_qwk_Task_Response": 0.5876268148422241,
+    "pred_variance_Task_Response": 6.448593750000001,
+    "val_qwk_Coherence_and_Cohesion": 0.5729712247848511,
+    "pred_variance_Coherence_and_Cohesion": 7.209999999999999,
+    "val_qwk_Lexical_Resource": 0.5792825222015381,
+    "pred_variance_Lexical_Resource": 6.469374999999999,
+    "val_qwk_Grammatical_Range_and_Accuracy": 0.5763903260231018,
+    "pred_variance_Grammatical_Range_and_Accuracy": 6.7435937500000005
+  },
+  "baseline_reference": {
+    "v2_4_overall_qwk": 0.7501,
+    "v2_4_adjacent_pct": 84.12,
+    "v2_4_exact_pct": 26.09,
+    "v2_4_rmse": 1.0084
+  }
+}

model_config.json

@@ -0,0 +1,18 @@
+{
+  "version": 2.5,
+  "config": {
+    "preset": "deberta_v3_base_en",
+    "sequence_length": 512,
+    "dropout": 0.08
+  },
+  "weights_file": "ielts_v2.5_base_en_10ep.weights.h5",
+  "calibration_file": "ielts_v2.5_base_en_10ep_calibration.pkl",
+  "bias_correction_offsets": [0.0, 0.0, 0.0, 0.0],
+  "serving": {
+    "variant": "augmented_calibrated",
+    "input_format": "essay_only",
+    "calibration_enabled": true,
+    "hf_repo": "koecheup/evalguide-ielts-v2.5"
+  },
+  "_note": "Production v2.5: augmented retrain + isotonic calibration. Weights are gitignored; download from HF or train locally. Rollback: set weights_file to ielts_v2.4_base_en_10ep.weights.h5 and version 2.4."
+}