korshunov.ai
English
Source · Hugging Face Forums
↑ Back to top of feed
media Hugging Face Forums · 2d ago

Buddy System: Rust entropy monitor with NER-gated uncertainty for tiered LLM inference

The Buddy System uses a Rust entropy monitor to detect per-token uncertainty in local Gemma 3 4B inference, routing only uncertain tokens to Sonnet via NER-gated span extraction and semantic retrieval. Benchmarks show it achieves 71.4% accuracy at $0.21, outperforming the Anthropic Advisor pattern (62.9% at $0.44) across seven Hugging Face datasets, with a key improvement on SQuAD v2 by routing source passage chunks to the cloud model.
media Hugging Face Forums · 3d ago

I built a novel triple-hybrid LLM under 1B parameters for ~$50

Mateusz has developed a full pre-trained language model, Project Inkblot's Titan v1, combining Mamba SSM, Multi-Head Attention, and 32-expert MoE in a single decoder-only architecture under 1B parameters. The model, trained on a single NVIDIA L4 GPU for ~$50, achieves 27.5 validation perplexity and demonstrates efficient scaling via a single-line config update, with all components implemented from scratch in PyTorch. Titan v2's first training cycle is now complete, and dataset expansion is underway.
media Hugging Face Forums · 3d ago

Seeking Indic Document Datasets for AI/OCR Training in India

QuantVectors is seeking annotated document datasets in Indic languages from India, including Hindi, Marathi, Gujarati, Bengali, Punjabi, Tamil, Urdu, Telugu, Odia, Kannada, Malayalam, and Assamese. The datasets must include invoice, receipt, utility bill, payment advice, packing list, commercial invoice, and credit note types, with approximately 400 documents per language, human-verified annotations, and 99%+ accuracy. Datasets must be commercially licensable and can be open-source or commercial, with a request for HuggingFace datasets, research datasets, or vendors specializing in this space.
media Hugging Face Forums · 3d ago

ML Surrogate Models in CFD/FEA: Real-World Practices and Challenges

Engineering practitioners report that graph neural networks and MLPs on parameterized designs offer the best practical balance for predicting fields like temperature and stress. Data efficiency is achievable with 10–50 training samples, especially when transfer learning is applied across similar geometries. Physics-informed neural networks (PINNs) remain largely experimental for complex engineering geometries, with most users relying on data-driven surrogates. Generalization remains a key challenge, with models often failing on out-of-distribution boundary conditions, prompting a return to full solver runs.
media Hugging Face Forums · 3d ago

Capability Is Not in the Weights: Empirical Negative Result on MLP Weight Projection

An empirical study found that projecting MLP weights from one transformer model into another fails to transfer semantic capability. Every tested variant performed worse than the unmodified host model, indicating a structural limitation in weight projection. The results challenge public claims about model capabilities based on benchmarks, showing such claims do not reflect actual internal weight geometry.
media Hugging Face Forums · 3d ago

The Clockwork Dark: A Local-First AI Narrative-RPG Engine

The Clockwork Dark is a local-first, AI-driven narrative-RPG engine that uses a deterministic state machine to resolve all game mechanics. It features two autonomous LLMs that narrate the story, with one acting as a patient world voice and the other as an unreliable, godlike assistant. The game offers players a choice: fight the encroaching supernatural corruption or embrace a quiet life in a bakery, with both paths considered valid endings.
media Hugging Face Forums · 3d ago

Small-scale debug comparison of OLMo-core with Engram graft

A 200-step training comparison between a base OLMo3 600M model and a version with a DeepSeek-style Engram graft shows lower training and evaluation loss, faster grad-norm stabilization, and improved early learning behavior. The Engram graft, injected into layers 1 and 5, increases trainable parameters to ~1.7B but maintains only a 40k increase in active parameters per token, indicating efficient memory usage.
media Hugging Face Forums · 3d ago

Hugging Face Spaces Proxy Strips CORS Credentials Header on OPTIONS Preflight

Users report that Hugging Face Spaces' edge proxy is now stripping the Access-Control-Allow-Credentials header from OPTIONS preflight requests, even when explicitly configured in Express middleware. The header is missing in responses despite code that sets it and includes a wildcard middleware for OPTIONS requests, suggesting the proxy intercepts and modifies the response before it reaches the backend container.