Saidzoda-Tajik-E5-Large
A Tajik text embedding model fine-tuned from multilingual-e5-large-instruct, used for Tajik document retrieval. The model supports inference on a single consumer-grade GPU, making Tajik retrieval available without large-scale infrastructure.
Timeline (last 90 days)
Saidzoda-Tajik-E5-Large is a Tajik text embedding model built on multilingual-e5-large-instruct (560M parameters).
Before Saidzoda-Tajik-E5-Large was released, there was no publicly described Tajik retrieval model.
The 560M-parameter model can be run for inference on a single consumer-grade GPU.
The training data contains 235,140 question-passage pairs, and training was run for one epoch.
During training, another embedding model was used to select hard negatives, and English and Russian question-passage pairs were mixed in to prevent the model from forgetting these two languages.
On legal texts the model has never seen, 93.6% of questions rank the correct passage in the top ten, compared with 80.2% for the baseline.
On tax documents, 94.5% of questions rank the correct passage in the top ten, compared with 82.6% for the baseline.
On the Belebele test set, Saidzoda-Tajik-E5-Large's top-10 retrieval accuracy improved from 92.2% for the baseline to 95.2%, but the confidence intervals overlap.
On legal and tax questions in the same style as the training problems, Saidzoda-Tajik-E5-Large ties with Giga-Embeddings-3B.
Training with bfloat16 caused the model to collapse.
1 more timeline events
All Facts (11)
Saidzoda-Tajik-E5-Large is a Tajik text embedding model built on multilingual-e5-large-instruct (560M parameters).
50%UnverifiedBefore Saidzoda-Tajik-E5-Large was released, there was no publicly described Tajik retrieval model.
50%UnverifiedThe 560M-parameter model can be run for inference on a single consumer-grade GPU.
50%UnverifiedThe training data contains 235,140 question-passage pairs, and training was run for one epoch.
50%UnverifiedDuring training, another embedding model was used to select hard negatives, and English and Russian question-passage pairs were mixed in to prevent the model from forgetting these two languages.
50%UnverifiedOn legal texts the model has never seen, 93.6% of questions rank the correct passage in the top ten, compared with 80.2% for the baseline.
50%UnverifiedOn tax documents, 94.5% of questions rank the correct passage in the top ten, compared with 82.6% for the baseline.
50%UnverifiedOn the Belebele test set, Saidzoda-Tajik-E5-Large's top-10 retrieval accuracy improved from 92.2% for the baseline to 95.2%, but the confidence intervals overlap.
50%UnverifiedOn legal and tax questions in the same style as the training problems, Saidzoda-Tajik-E5-Large ties with Giga-Embeddings-3B.
50%UnverifiedTraining with bfloat16 caused the model to collapse.
50%UnverifiedThe legal, tax and Wikipedia test questions come from the same model that generated the training questions, so these scores should be regarded as an upper bound.
50%