Serovvans/trocr-prereform-orthography
61.6M • Updated • 15 • 4
Boris Orekhov
nevmenandr
AI & ML interests
Natural Language Processing, Poetry Generation, Linguistics, Low-resource languages
Recent Activity
posted an update about 17 hours ago
https://huggingface.co/nevmenandr/char-based-lstm-russian-poetry-pasternak
🧠 LSTM Language Model Visualization: A Deep Dive into Char-RNN
📊 Model Architecture at a Glance
- Model Type: 5-layer LSTM
- Hidden Size: 512
- Vocabulary: 137 characters
- Sequence Length: 50
- Total Parameters: ~9.8 million
- Training: 50 epochs, 10,750 iterations
- Final Validation Loss: 1.1266
- The model learned to generate Pasternak-style poetry - pretty impressive for a char-rnn!
🎨 The Beautiful Mess
Check out this heatmap visualization - it's like a Persian carpet! 🏠✨
- Each gate has its own patterns:
- Input Gate: Controls what new info enters the cell
- Forget Gate: Decides what to discard
- Cell Gate: Creates new candidate values
- Output Gate: Determines what to output
- The weights show beautiful structured patterns - different gates learned distinct strategies for processing
text.https://huggingface.co/papers/2306.02771 posted an update 10 days ago
🔥 New Russian Stylometry Dataset!
Russian Stylometric Dataset (RSD) — 322 texts from the 19th – early 20th centuries (16 million words), prepared for analysis in stylo (R) and machine learning (Python).
📚 What's inside?
Fiction, journalism, scientific texts, drama, poetry
Grouped by author, gender, age, genre, literary movements (Romanticism/Realism)
Character speech (Tolstoy, Gogol, Ostrovsky)
Generated texts (LSTM, GPT)
📊 Use cases: authorship attribution, clustering, classification, benchmarking methods.
🔓 Public domain + GPL-3.0 license.
👉 Learn more: https://github.com/nevmenandr/RSD
DOI: 10.5281/zenodo.20701309 posted an update 12 days ago
https://huggingface.co/nevmenandr/char-based-lstm-russian-poetry-https://huggingface.co/nevmenandr/char-based-lstm-russian-poetry-mandelshtam
https://huggingface.co/nevmenandr/char-based-lstm-russian-poetry-hexameter
https://huggingface.co/papers/2306.02771
📜 RNN vs. Transformers: How an Old Architecture Better Perceives Poetic Style
In the era of Transformer dominance, we often forget that old RNNs (especially character-level LSTMs) remain irreplaceable for tasks where *individual style*, rhythm, and micro-patterns matter. These three models are clear proof of that.
🎯 Why does this matter today?
- **Stylistic analysis**: RNNs better capture meter, repetitions, and unexpected tonal shifts.
- **Teaching poetics**: generating "almost correct" but hallucinating lines helps explore the boundaries of style.
- **Nostalgia and replication**: a reminder that not everything is measured by BLEU and perplexity.
🖼️ Visualization
Attached is an infographic comparing the three models (architecture, style, generation sample).
> RNNs aren't dead. They're just writing poetry in silence.Organizations
liked a
model 10 months ago
liked 2
models about 2 years ago