ScaleDown
An AI company that released Constrained Decoding Models (CDM), focused on task-specific small models. Its models include Classify, Extract, Summarize, and Compress, and it has a partnership with LangChain.
Timeline (last 90 days)
ScaleDown released four small task models, called Constrained Decoding Models (CDM), each trained for a single task and only returning answers allowed by that task.
Among ScaleDown CDM's four models, Classify and Extract are pointer-type, Summarize is writing-type, and Compress does not clearly state whether it selects or rewrites.
ScaleDown named its own product Constrained Decoding Models, which is easily confused with the general constrained decoding technique, but the former is a specific vendor's task-specific small models, while the latter is a decoding-time constraint method usable with any LLM.
ScaleDown suggests surrounding the main LLM with small models: small models handle routine classification, extraction, and compression, while the general-purpose model handles the parts that require generation.
ScaleDown's Classify model always returns a winner even when the model is not certain, so in practice one should add an other label and test on one's own samples.
The boundaries of the pointer-style approach include: the answer must be in the input, a valid label is not the same as a correct label, Summarize is writing-type while Compress is not explicitly stated, the large numbers come from ScaleDown itself with no third-party verification, and text leaves the server.
All Facts (6)
Among ScaleDown CDM's four models, Classify and Extract are pointer-type, Summarize is writing-type, and Compress does not clearly state whether it selects or rewrites.
50%UnverifiedScaleDown named its own product Constrained Decoding Models, which is easily confused with the general constrained decoding technique, but the former is a specific vendor's task-specific small models, while the latter is a decoding-time constraint method usable with any LLM.
50%UnverifiedScaleDown suggests surrounding the main LLM with small models: small models handle routine classification, extraction, and compression, while the general-purpose model handles the parts that require generation.
50%UnverifiedScaleDown's Classify model always returns a winner even when the model is not certain, so in practice one should add an other label and test on one's own samples.
50%UnverifiedThe boundaries of the pointer-style approach include: the answer must be in the input, a valid label is not the same as a correct label, Summarize is writing-type while Compress is not explicitly stated, the large numbers come from ScaleDown itself with no third-party verification, and text leaves the server.
50%UnverifiedScaleDown released four small task models, called Constrained Decoding Models (CDM), each trained for a single task and only returning answers allowed by that task.
50%