Multi-task classifier fine-tuned tu BamiBERT cho van ban vaccine tieng Viet, 3 head:
misinfo: {Fake, Real}
stance: {Favor, Against, Neutral}
sentiment: {Negative, Neutral, Positive}
Input: raw text (BamiBERT byte-BPE, KHONG word-segmentation), max_len 256. Train 10 epoch, class-weighted loss.
Cung du lieu + split (seed=42) voi PhoBERT-v2 de so cong bang.
So sanh voi PhoBERT-v2 (cung eval harness, benchmark_test_set_v3, 186 mau)
Model
misinfo macro-F1
Fake-recall
stance
sentiment
PhoBERT-v2 (argmax)
0.6941
0.6071
0.6570
0.7279
BamiBERT (argmax 0.5)
0.7458
0.5000
0.6423
0.7446
BamiBERT (thr=0.325, val-chon)
0.7688
0.6071
0.6423
0.7446
Ket luan: BamiBERT + hieu chinh nguong (chon tren VAL seed=42, ap len test, zero-leakage) dat
misinfo macro-F1 0.7688 (+0.075 so PhoBERT) voi Fake-recall bang PhoBERT (0.6071) - KHONG can re-train.
Theo nguyen tac "model quyet nhan, code quyet nguong/do tin". Diem manh phu: BamiBERT bo word-segmentation
(raw text) + context 2050 vs PhoBERT 256.
Han che
Test nho (28 Fake / 186 mau). Y nghia thong ke cho benchmark vaccine-VN lon hon.
Nguong 0.325 chon tren 10% validation split (seed 42). Metric tu lan chay tham chieu (train co randomness ngoai split).
