tamizhimet-xlmr-base

FacebookAI/xlm-roberta-base fine-tuned to classify a Tamil song-lyric line into one of 9 rhetorical-figure classes (TamizhiMet).

Labels: Literal, Uvamai, Uruvagam, Uyarvu_Navirchi_Ani, Thatkurippetra_Ani, Vanchapugalchi_Ani, Madakku, Ilivu, Iyam.

Training

• Data: TamizhiMet train split, 8,680 lines (one label per line).
• Epochs 5, batch size 32, learning rate 3e-05, AdamW (weight decay 0.01), linear schedule with 10% warmup, max length 64 tokens.
• Cross-entropy with class weights proportional to n_c^-0.5 (mean 1).
• Seed 42; precision fp16; checkpoint after the last epoch.

Results of this checkpoint

Macro-F1 is averaged over the classes that have gold lines in the split.

Per-class F1 (gold support in brackets; fewer than 10 gold lines is unreliable):

For reference, the paper runs (same settings) score macro-F1 val 0.349 ± 0.011; test 0.291 ± 0.008 (mean ± sd over seeds 42/43/44). This checkpoint is a retrained seed, so its scores differ slightly.

Limitations

The training data is dominated by Literal and Uvamai; the rarest classes have very few training lines, and the model seldom or never predicts them (see the per-class table).

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo = "kavienanj/tamizhimet-xlmr-base"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo).eval()

line = "பறவை போல் பறந்தவன்"
enc = tok(line.strip() + '', add_special_tokens=True, truncation=True, max_length=64, return_tensors="pt")
with torch.no_grad():
    probs = model(**enc).logits.softmax(-1)[0]
print(model.config.id2label[int(probs.argmax())])