Update README.md

README.md

@@ -1,3 +1,77 @@
----
-license: mit
----
+---
+license: mit
+---
+
+
+
+## Training data
+* **Source:** custom, manually annotated CBDC sentences
+* **Size:** 2,405 sentences
+
+**Class distribution:**
+* `neutral`: 1,068 (44.41%)
+* `positive`: 1,026 (42.66%)
+* `negative`: 311 (12.93%)
+
+**Splits** (row-wise, stratified by label):
+* **train:** 1,924
+* **validation:** 240
+* **test:** 241
+
+---
+
+## Preprocessing
+
+* Lowercased, raw sentences (no stemming or lemmatization)
+* Tokenization: base model tokenizer (`bilalzafar/cb-bert-mlm`), **max\_length=320**, truncation enabled
+* Dynamic padding via `DataCollatorWithPadding`
+
+---
+
+## Training procedure
+
+* **Base model:** `bilalzafar/cb-bert-mlm`
+* **Head:** `AutoModelForSequenceClassification` with 3 labels
+* **Optimizer:** AdamW (via HF Trainer)
+* **Learning rate:** 2e-5
+* **Batch size:** 16 (train/eval)
+* **Epochs:** up to 8 with early stopping (patience=2); best epoch \~6
+* **Warmup ratio:** 0.06
+* **Weight decay:** 0.01
+* **Precision:** fp16
+* **Seed:** 42
+* **Hardware:** Google Colab (T4)
+
+---
+
+## Class imbalance & loss
+
+* **Loss:** Focal Loss with γ = 1.0
+* **Class weights:** computed from the **train split** (`class_weight="balanced"`) and applied in the loss
+* **Sampler:** `WeightedRandomSampler` with √(inverse frequency) per-sample weights
+
+---
+
+## Evaluation
+
+**Validation** (\~10% split):
+* accuracy: **0.8458**
+* macro-F1: **0.8270**
+* weighted-F1: **0.8453**
+
+**Test** (\~10% split):
+accuracy: **0.8216**
+macro-F1: **0.8121**
+weighted-F1: **0.8216**
+
+**Per-class (test):**
+
+| class    | precision | recall | f1     | support |
+| -------- | --------- | ------ | ------ | ------- |
+| negative | 0.8214    | 0.7419 | 0.7797 | 31      |
+| neutral  | 0.7857    | 0.8224 | 0.8037 | 107     |
+| positive | 0.8614    | 0.8447 | 0.8529 | 103     |
+
+> Note: On the **entire annotated set** (in-domain evaluation, not a hold-out),
+> the same model reaches \~0.95 accuracy / weighted-F1.
+> Treat those as upper bounds; the **test split** above is the recommended reference.