DSpark and DFlash2 draft models for faster inference with speculative decoding
Yosef Worku Alemneh
rasyosef
AI & ML interests
Pretraining, Supervised Fine Tuning, Direct Preference Optimization, Retrieval Augmented Generation (RAG), Function Calling
Recent Activity
updated a model about 3 hours ago
rasyosef/Llama-3.2-3B-Instruct-DSpark updated a model about 13 hours ago
rasyosef/Qwen3.5-2B-DSpark updated a dataset about 17 hours ago
rasyosef/open-perfectblend-qwen-3.5-2bOrganizations
SPLADE sparse retrieval models based on BERT-Tiny (4M) and BERT-Mini (11M) distilled from a Cross-Encoder on the MSMARCO dataset
Llama 3.2 decoder transformer models trained on Amharic text
-
rasyosef/Llama-3.2-400M-Amharic
Text Generation • 0.4B • Updated • 54 • 3 -
rasyosef/Llama-3.2-400M-Amharic-Instruct-Poems-Stories-Wikipedia
Text Generation • 0.4B • Updated • 56 • 1 -
rasyosef/Llama-3.2-400M-Amharic-Instruct
Text Generation • 0.4B • Updated • 181 -
rasyosef/Llama-3.2-180M-Amharic
Text Generation • 0.2B • Updated • 208 • 1
GPT2 transformer decoder models pretrained on 290 million tokens of Amharic text
These models underwent supervised fine-tuning and direct preference optimization for instruction following on top of Microsoft's Phi 2 base LLM
Amharic Neural Information Retrieval models: Bi-Encoders, Cross Encoders, ColBERT, and SPLADE
Text Embedding and ColBERT models based on Amharic RoBERTa and BERT for Amharic passage retrieval
-
Optimized Text Embedding Models and Benchmarks for Amharic Passage Retrieval
Paper • 2505.19356 • Published -
rasyosef/roberta-amharic-text-embedding-base
Sentence Similarity • 0.1B • Updated • 1.14k -
rasyosef/colbert-roberta-amharic-base
Sentence Similarity • 0.1B • Updated • 28 -
rasyosef/roberta-amharic-text-embedding-medium
Sentence Similarity • 42.1M • Updated • 151
BERT and RoBERTa transformer encoder models pretrained on 290 million tokens of Amharic text
These models underwent supervised fine-tuning and direct preference optimization for instruction following on top of Microsoft's Phi 1.5 base LLM
Instruction-tuned (chat) versions of Nvidia's Minitron base models created through supervised fine-tuning (SFT)
DSpark and DFlash2 draft models for faster inference with speculative decoding
Amharic Neural Information Retrieval models: Bi-Encoders, Cross Encoders, ColBERT, and SPLADE
SPLADE sparse retrieval models based on BERT-Tiny (4M) and BERT-Mini (11M) distilled from a Cross-Encoder on the MSMARCO dataset
Text Embedding and ColBERT models based on Amharic RoBERTa and BERT for Amharic passage retrieval
-
Optimized Text Embedding Models and Benchmarks for Amharic Passage Retrieval
Paper • 2505.19356 • Published -
rasyosef/roberta-amharic-text-embedding-base
Sentence Similarity • 0.1B • Updated • 1.14k -
rasyosef/colbert-roberta-amharic-base
Sentence Similarity • 0.1B • Updated • 28 -
rasyosef/roberta-amharic-text-embedding-medium
Sentence Similarity • 42.1M • Updated • 151
Llama 3.2 decoder transformer models trained on Amharic text
-
rasyosef/Llama-3.2-400M-Amharic
Text Generation • 0.4B • Updated • 54 • 3 -
rasyosef/Llama-3.2-400M-Amharic-Instruct-Poems-Stories-Wikipedia
Text Generation • 0.4B • Updated • 56 • 1 -
rasyosef/Llama-3.2-400M-Amharic-Instruct
Text Generation • 0.4B • Updated • 181 -
rasyosef/Llama-3.2-180M-Amharic
Text Generation • 0.2B • Updated • 208 • 1
BERT and RoBERTa transformer encoder models pretrained on 290 million tokens of Amharic text
GPT2 transformer decoder models pretrained on 290 million tokens of Amharic text
These models underwent supervised fine-tuning and direct preference optimization for instruction following on top of Microsoft's Phi 1.5 base LLM
These models underwent supervised fine-tuning and direct preference optimization for instruction following on top of Microsoft's Phi 2 base LLM
Instruction-tuned (chat) versions of Nvidia's Minitron base models created through supervised fine-tuning (SFT)