2 days ago
-
Upload modeling_maria.py with huggingface_hub
-
Upload modeling_maria.py with huggingface_hub
-
Upload modeling_maria.py with huggingface_hub
-
Upload README.md with huggingface_hub
-
Upload README.md with huggingface_hub
-
Upload README.md with huggingface_hub
-
Upload README.md with huggingface_hub
-
Upload README.md with huggingface_hub
-
Upload README.md with huggingface_hub
-
Upload README.md with huggingface_hub
-
Upload generation_config.json with huggingface_hub
-
Upload modeling_maria.py with huggingface_hub
-
Upload config.json with huggingface_hub
-
Fix _tied_weights_keys to mapping dict and use fused SDPA causal attention
-
Restore exact trained Fused SDPA attention implementation from pretrain codebase
-
Update quickstart snippet with pad_token_id and fix character encoding
-
Include lm_head.weight explicitly in safetensors checkpoint (363 tensors, zero missing weights)
-
Add GenerationMixin and tie_weights(**kwargs) support
-
Set tie_word_embeddings: true
-
Universal import for MariaConfig (supports both package and standalone local runs)
-
Inherit from GenerationMixin for full transformers 4.50+ generation compatibility
-
Use relative import from .configuration_maria for HF dynamic module compatibility
-
Embed crisp 4K architecture.png and remove unrendered Mermaid block
-
Add 4K publication-grade architecture.png diagram
-
Replace clunky monospace ASCII block with high-res SVG and interactive Mermaid diagram
-
Add publication-grade dark vector graphic architecture.svg
-
Inherit from GenerationMixin in MariaForCausalLM
-
Fix self.num_hidden_layers attribute typo in MariaConfig
-
Upgrade model card to publication-grade colorful visual aesthetic with shields badges and callouts
-
Clarify architectural context: HyperGQE preprint (TinyLlama-1.1B testbed) vs Maria-1B (from-scratch 20B token pretraining)
-
Add authentic research paper preprint DOI citation (Zenodo: 10.5281/zenodo.22747827)
-
Fix kwargs pop in MariaConfig for tie_word_embeddings
-
Initial release: Maria-1B 20B tokens pretrained model with HyperGQE
-
initial commit
