- Major Update : Allowed Online contrastive loss in model training with paired data with labels - Fixed the behaviour of the custom suffix feature in merge functions - Fixed an incorrect piece of preprocessing code for paired data with labels - we recommend training such models again for a substantial increase in performance - Allowed loss_type in the train_model args - "supcon" and "onlinecontrastive" - Made some changes in the linkage configs to allow loss params into the training args. loss_params is a dictionary containing "temperature" for supcon loss and "margin" for onlinecontrastive.
0.1.12
- Fixes in merge and classification inference to be in line with OpenAI API changes - Bug fixes in column serialisation when using open ai embeddings - Updated toml file to restrict by package versioning - forward compatibility would not be supported - Made merge with blocking faster - no longer needs loading the model for each block - All Linkage inference functions can now take in a LinkTransformer model (wrapper around SentenceTransformer) as input. This would be useful for workflows requiring looping; preventing repeated model loading. - More robust typing where it was lacking - Fixes in cluster functions and bringing them to the root directory
0.1.11
- Allowed a progress bar for training inference - Fixed a bug in tokenizer saving
0.1.10
- Allowed a progress bar for classification inference
0.1.9
- Minor bug fixes
0.1.8
- More fixes; performance improvements - Added classification features - infer using HF transformers, OpenAI (Chat), and train custom transformer models in 1 line - More demo datasets