jack1211son/abdta-multi-task-v4-0pct-2x-equal

Multi-task tone and aspect: 0%_2x_equal

Final checkpoint in the aspect-based risk-disclosure tone analysis (ABDTA) collection for Indonesian listed companies' annual-report narratives. The collection is intended for final models retrained on training plus validation data. The original downloaded folder is model_multi_task_v4_0%_2x_equal; the learned weights are unchanged.

0pct / 25pct in the repository name means training-time aspect-keyword dropout. 2x / 4x denote pseudo-label augmentation variants; equal balances aspects and train-dist follows the training distribution. No augmentation suffix means baseline.

Tone label order: ['Positive', 'Negative', 'Neutral'].

Aspect label order: ['Economics', 'Profit|Loss', 'Dividend', 'Financing', 'Brand', 'Technology', 'Investment', 'Governance', 'Product|Service', 'M&A', 'Legal', 'Environment', 'Social&People', 'Political'].

Use the same risk and aspect keyword extraction as the project notebooks. Input format:

[Risk Keywords: <risk words> Aspect Keywords: <aspect words>] <sentence>

Inputs are limited to 256 tokens. This checkpoint classifies disclosure language; it does not directly predict stock prices or investment returns.

Load

import torch
from transformers import AutoTokenizer, AutoModel

repo = "jack1211son/abdta-multi-task-v4-0pct-2x-equal"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModel.from_pretrained(repo, trust_remote_code=True).eval()
text = "[Risk Keywords: increased Aspect Keywords: revenue] Revenue increased during the year."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    outputs = model(**inputs)
logits = outputs["tone_logits"], outputs["aspect_logits"]

Multi-task models use their bundled custom architecture code. The package contains weights, configuration, tokenizer assets, and this model card; source sentence-level evaluation CSVs are not published as model assets.