Ferman slot tagger (XLM-R base, v2)
Tags the slots (date, time, event name, list name, person, place and others)
inside assistant commands in English, Arabic, Sorani and Badini, using BIO
labels. Fine-tuned from xlm-roberta-base. Pairs with
OmarSY11/xlmr-intent-v2. Part
of the Ferman project: https://github.com/omarGH99/ferman-ai-assistant
Usage
from transformers import pipeline
tag = pipeline("token-classification", model="OmarSY11/xlmr-slot-v2", aggregation_strategy="simple")
print(tag("remind me to call mum tomorrow at 7pm"))
Training data
Rows of the v2 configuration of OmarSY11/ferman-nlu
that carry slot annotations (annot column), train split. English and Arabic
derive from MASSIVE; Sorani and Badini are machine-translated.
Arabic caveat: only 304 Arabic rows carry slot annotations in this release (MASSIVE has more; they were not carried over), so Arabic slot extraction is barely trained and was not measured.
Training
| Base model | xlm-roberta-base |
| Max length | 64 tokens |
| Learning rate / batch size | 2e-5 / 32 |
| Epochs | up to 8, early stopping (patience 2) on validation F1 |
| Warm-up / weight decay | 0.1 / 0.01 |
| Low-resource handling | Sorani, Badini and Arabic training rows duplicated once |
| Seed | 42 |
The notebook is in the repo: ml/train_xlmr_slot_v2_runpod.ipynb.
Results
Entity-level (seqeval) on the held-out test split of the v2 data.
| Language | Test rows scored | Precision | Recall | F1 |
|---|---|---|---|---|
| English | 1,954 | 0.790 | 0.844 | 0.816 |
| Sorani | 370 | 0.620 | 0.669 | 0.643 |
| Badini | 370 | 0.598 | 0.649 | 0.623 |
| Arabic | 0 | not measured | not measured | not measured |
| Overall | 0.743 | 0.796 | 0.769 |
Kurdish test sets are small, so those figures carry wide uncertainty.
Limitations
- Arabic slots are unmeasured and trained on very little data.
- Kurdish data is machine-translated and informally reviewed by one person (native Badini, non-native Sorani).
- Slots come from a fixed MASSIVE-style label set; unseen slot types are not handled.
License and citation
Model weights: MIT, like the base model. Training data: CC BY 4.0, derived from MASSIVE (FitzGerald et al., 2022, arXiv:2204.08582), which must be credited. Please also cite https://github.com/omarGH99/ferman-ai-assistant.