You can deploy this model and start earning money today!

Fakhri2503/
FineTuneBi-xLSTM

Name: FineTuneBi-xLSTM
Availability: InStock

Views

No views yet

Bi-xLSTM-CRF e2e ABSA (HOASA)

Fine-tuned ELMo-style Bi-xLSTM (Fakhri2503/xLSTM) untuk NER task pada dataset HOASA.

Arsitektur

Encoder : Bi-xLSTM ELMo-style pretrained (Fakhri2503/xLSTM)
Head : LayerNorm → Linear(1536→768) → GELU → Linear(768→num_labels)
Decoder : BIOES-constrained CRF

Preprocessing

Lowercase, NFKC normalization, emoji removal
First-subword label alignment
Class-weighted CE loss + CRF loss

Hasil Ablation (Test Set)

Strategy	Precision	Recall	F1
Full Freeze	0.7571	0.6830	0.7181
Full Unfreeze	0.7848	0.7299	0.7563

Best: Full Unfreeze (F1=0.7563)

Label Format

BIOES format, 12 labels: O, B-NEG, B-NEU, B-POS, E-NEG, E-NEU, E-POS, I-NEG, I-POS, S-NEG, S-NEU, S-POS

Usage

python

# Load config
from huggingface_hub import hf_hub_download
import json, torch

config_path = hf_hub_download('Fakhri2503/FineTuneBi-xLSTM', 'config.json')
with open(config_path) as f: config = json.load(f)
label2id = config['label2id']
id2label = {int(k): v for k, v in config['id2label'].items()}

# Load model (perlu class BiXLSTMLM + BiXLSTMCrfNER dari notebook)
ckpt = hf_hub_download('Fakhri2503/FineTuneBi-xLSTM', 'full_unfreeze_model.pt')
model.load_state_dict(torch.load(ckpt, map_location='cpu'))