alto-tfgbv-all

A multilingual binary TFGBV classifier for Arabic, Swahili, French fine-tuned on afro-xlmr-large-76L with a three-output sigmoid head (gendered_content, simple_hate, tfgbv); the TFGBV label is the third output ≥ 0.5. Training data: Bench-train + ALTO-train combined, 10 epochs, three sigmoid heads trained jointly from meedan/alto config all (export 20260924), using the unredacted doxxing text for the items that are redacted in the public dataset.

Training

  • Epochs: 10
  • LR: 5e-06
  • Batch: 32
  • Max_length: 128
  • Seed: 42
  • 10% linear warmup
  • class-weighted BCE per head (pos_weight = n_neg/n_pos of each label), head losses summed
  • Training items: {"alto": 1693, "bench": 7512}

TFGBV-class F1 on the ALTO gold test split, per language, this checkpoint vs. the zero-shot baselines of the ALTO paper (same items, same labels):

model TFGBV F1 (ar, n=117) TFGBV F1 (fr, n=180) TFGBV F1 (sw, n=127)
AfroXLMR-L-76L (M) (this model) 0.772 0.700 0.672
OSS-SG-120B 0.764 0.624 0.526
OSS-SG-20B 0.761 0.652 0.600
LFTW 0.710 0.589 0.463
P-API 0.745 0.600 —

P-API (Perspective) does not support Swahili.

Usage

The weights are a plain safetensors state dict for the MultiLabelClassifier module below (shared encoder, one Linear(hidden, 3) head). The three sigmoid outputs are, in order, gendered_content, simple_hate and tfgbv; the TFGBV label is tfgbv_prob >= 0.5 (the first two heads are auxiliary signals, not part of the decision).

import torch, torch.nn as nn
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from transformers import AutoModel, AutoTokenizer

BASE, REPO = "Davlan/afro-xlmr-large-76L", "meedan/alto-tfgbv-xlmr-large-3L"
LABELS = ["gendered_content", "simple_hate", "tfgbv"]

class MultiLabelClassifier(nn.Module):
    def __init__(self, model_id, n_labels=3):
        super().__init__()
        self.backbone = AutoModel.from_pretrained(model_id)
        self.dropout = nn.Dropout(0.1)
        self.head = nn.Linear(self.backbone.config.hidden_size, n_labels)
    def forward(self, input_ids, attention_mask):
        out = self.backbone(input_ids=input_ids, attention_mask=attention_mask)
        pooled = out.pooler_output if getattr(out, "pooler_output", None) is not None else out.last_hidden_state[:, 0]
        return self.head(self.dropout(pooled))

tok = AutoTokenizer.from_pretrained(BASE)
model = MultiLabelClassifier(BASE).eval()
model.load_state_dict(load_file(hf_hub_download(REPO, "model.safetensors")))

@torch.no_grad()
def predict(texts, max_length=128):
    enc = tok(texts, truncation=True, padding="max_length", max_length=max_length, return_tensors="pt")
    probs = torch.sigmoid(model(enc["input_ids"], enc["attention_mask"]))
    return [dict(zip(LABELS, p.tolist())) | {"tfgbv_label": int(p[2] >= 0.5)} for p in probs]
Downloads last month

-

Downloads are not tracked for this model. How to track
Safetensors
Model size
0.6B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for meedan/alto-tfgbv-xlmr-large-3L

Finetuned
(19)
this model

Dataset used to train meedan/alto-tfgbv-xlmr-large-3L

Collection including meedan/alto-tfgbv-xlmr-large-3L