TinyModel1

TinyModel1 is a compact encoder model for news topic classification, trained from scratch on the AG News dataset. It targets fast CPU/GPU inference, simple deployment behind a router or API, and use as a baseline before larger or domain-specific models.

Links

• Source code (train & export): https://github.com/HyperlinksSpace/TinyModel
• Live demo (Space): TinyModel1Space (canonical Hub URL; avoids unreliable *.hf.space links)

---

Model summary

Task	Text classification (single-label, 4 classes)
Labels	World, Sports, Business, Sci/Tech
Architecture	Tiny BERT-style encoder (BertForSequenceClassification)
Parameters	1,339,268 (~1.34M)
Max sequence length	128 tokens (training & inference)
Framework	Transformers · Safetensors

---

Model overview

This release fits a small footprint so you can run batch or interactive classification without heavy GPUs. Training uses a WordPiece tokenizer fit on the training split and a shallow BERT stack suited to short news sentences.

Core capabilities

• Topic routing — assign one of four coarse news categories for search, feeds, or moderation triage.
• Low latency — small parameter count keeps inference suitable for edge and serverless setups.
• Fine-tuning base — swap labels or add data for your domain while keeping the same architecture.

---

Training

Setting	Value
Train samples	3000
Eval samples	600
Epochs	2
Batch size	16
Learning rate	0.0001
Optimizer	AdamW

---

Evaluation

Metric	Value
Eval accuracy	0.5150
Final train loss	1.1568

Metrics are computed on the held-out eval split configured above; treat them as a sanity-check baseline, not a production SLA.

---

Getting started

Inference with transformers

from transformers import pipeline

clf = pipeline(
    "text-classification",
    model="TinyModel1",  # or local path after save
    tokenizer="TinyModel1",
    top_k=None,
)
text = "Markets rose after the central bank held rates steady."
print(clf(text))

Use top_k=None (or your Transformers version’s equivalent) to obtain scores for all labels. Replace "TinyModel1" with your Hugging Face model id (for example HyperlinksSpace/TinyModel1) when loading from the Hub.

---

Training data

• Dataset: fancyzhx/ag_news (4-class news topics).
• Preprocessing: tokenizer trained on training texts; sequences truncated to 128 tokens.

---

Intended use

• Prototyping routing, tagging, and dashboard features over English news-style text.
• Teaching and benchmarking small-classification setups.
• Starting point for domain adaptation (finance, sports, etc.) with your own labels.

---

Limitations

• Accuracy is modest by design; do not rely on it for high-stakes decisions without validation on your data.
• English-oriented news wording; other languages or social-style text may degrade.
• Four fixed classes; not suitable as a general-purpose language model.

---

License

This model is released under the Apache 2.0 license (see repository LICENSE where applicable).