Reza2kn/Hes-Shenaas-RizehPizeh-v0.1

Hes Shenaas RizehPizeh v0.1

RizehPizeh is the compact Hes Shenaas Persian emotion classifier: a Persian ALBERT encoder followed by a one-layer bidirectional GRU classification head. It contains 12,040,967 parameters and predicts seven labels.

Smallest Persian emotion classifier

As of September 9, 2026, to the best of our knowledge, this Hes Shenaas is the smallest publicly available Persian text emotion recognizer/classifier, with 12M parameters. This claim is based on a review of discoverable public model releases and published Persian emotion classifiers. It does not cover unpublished or unindexed models.

Label	Meaning
ANGRY	anger
FEAR	fear or anxiety
HAPPY	happiness or joy
HATE	hate, disgust, or strong aversion
SAD	sadness
SURPRISE	surprise
OTHER	neutral, unclear, or outside the six emotions

Quick start

This repository contains the encoder, GRU head, tokenizer, and model code. No second model repository is downloaded at inference time.

pip install "torch>=2.4" "transformers>=4.45"

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

repo = "Reza2kn/Hes-Shenaas-RizehPizeh-v0.1"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained(
    repo,
    trust_remote_code=True,
).eval()

text = "امروز واقعاً روز فوق‌العاده‌ای بود"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
with torch.inference_mode():
    probabilities = model(**inputs).logits.softmax(dim=-1)[0]

scores = {
    model.config.id2label[index]: float(score)
    for index, score in enumerate(probabilities)
}
print(max(scores, key=scores.get), scores)

Pipeline API:

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Reza2kn/Hes-Shenaas-RizehPizeh-v0.1",
    trust_remote_code=True,
)
print(classifier("از این وضعیت خیلی عصبانی‌ام", top_k=None))

Evaluation

The checkpoint was selected using a fixed 1,232-example validation split. The protected 1,151-example test split was evaluated once after selection.

Split	Examples	Accuracy	Macro F1	Weighted F1
Validation	1,232	72.65%	73.24%	72.46%
Test	1,151	70.03%	69.17%	70.20%

For scale, Gemini 3.8 Flash scored 78.11% accuracy (899/1,151) on the same frozen test set under the project's fixed prompting contract. RizehPizeh is 8.08 percentage points behind that hosted reference while remaining a self-contained local model with only 12,040,967 parameters.

Test F1 by class:

ANGRY	FEAR	HAPPY	HATE	SAD	SURPRISE	OTHER
69.23%	71.67%	72.88%	63.79%	76.21%	68.02%	62.37%

The packaged model was reloaded through the Transformers auto classes and its predictions were checked against all 1,151 stored frozen-test predictions.

Architecture and training

• Persian ALBERT encoder revision: 1a4861062a5501088ce06389f5d67921e47320e4
• Bidirectional GRU: 128 hidden units per direction
• Dropout: 0.2
• Linear classifier: 256 → 7
• Maximum input length: 128 tokens
• Selected epoch: 2
• Training data: 4,862 cleaned original rows plus 42,000 balanced public Persian tweets carrying high-confidence teacher probabilities
• Distillation threshold: 0.95, capped at 6,000 external rows per class
• Loss: original-label cross entropy plus temperature-2 teacher KL divergence
• Protected validation and test texts were excluded from external API payloads

Limitations

Emotion classification is subjective. Accuracy will vary on formal prose, sarcasm, code-switching, dialects, long documents, and domains unlike the benchmark. OTHER combines neutral and ambiguous cases. HATE includes strong aversion and disgust under this dataset's label contract. Do not use predictions as the sole basis for high-impact decisions about people.

Integrity

The original training checkpoint SHA-256 was:

9fc3e1d865d7177a0f2a3c5d90e2114357b355b98ac0b80292a7d9d5e9a69d38

The converted model.safetensors checksum is recorded in release.json.