YAML Metadata Warning:The pipeline tag "text2text-generation" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other

mT5-Large fine-tuned on Romanian ASI

Fine-tuned google/mt5-large (1.2B params) on the Romanian Affective State Identification (ASI) benchmark, following the MASIVE (Deas et al. 2024) recipe.

Task

Given a Romanian text with an affective-state word masked as <extra_id_0>, predict the masked word. Example:

Input:  Mă simt foarte <extra_id_0> după ce am terminat cursul.
Target: <extra_id_0> mândru <extra_id_1>

Training

Training data 45,181 (text, masked_word) pairs extracted by pattern matching + LLM validation + human eval from Filmot, FULG, and 6 small Romanian datasets
Optimizer Adafactor, lr 4e-4 linear decay, weight decay 0.01
Batch size 16, 3 epochs (8,472 steps)
Precision bf16
Hardware 1× NVIDIA RTX A6000 48 GB
Wall clock 2h 04m
Final val loss 0.350

Evaluation (beam-5)

Test n Acc@1 Acc@3 Acc@5 MRR
val (seen vocab) 2,658 57.7% 77.5% 83.2% 0.68
test (unseen vocab) 5,315 0.79% 1.52% 2.20% 0.013
zero-shot mT5-large baseline, unseen 5,315 14.5% 18.7% 18.8% 0.17

On unseen vocabulary, Sim@1 (contextual BERT cosine) = 0.74: the fine-tune predicts semantic near-synonyms of held-out emotion words (e.g. frică for rușine).

Full details in the repo: https://github.com/Continual-Learning-Emotion-Group/Romanian_ASI/tree/mt5-finetune/pipeline/ft_mt5

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tok = AutoTokenizer.from_pretrained("alexjerpelea/mt5-Large-Romania")
m   = AutoModelForSeq2SeqLM.from_pretrained("alexjerpelea/mt5-Large-Romania")

text = "Am fost foarte <extra_id_0> după ce am terminat proiectul."
ids = tok(text, return_tensors="pt").input_ids
out = m.generate(ids, num_beams=5, num_return_sequences=5, max_new_tokens=10)
for o in out:
    print(tok.decode(o, skip_special_tokens=False))
Downloads last month
316
Safetensors
Model size
1B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for alexjerpelea/mt5-Large-Romania

Base model

google/mt5-large
Finetuned
(96)
this model