emillykkegrann/edda-v0.2-GGML

Edda v0.2 GGML

GGML conversions of danish-foundation-models/edda-v0.2 at revision 98ec5eee028b111a19733890fe32ea65267b662b, for whisper.cpp-compatible runtimes. Edda is by the Alexandra Institute, trained within the CoRal project and released by Danish Foundation Models. All credit for the model belongs to them.

This repository exists so Codictate can offer Edda as a Danish speech model. It is not affiliated with or endorsed by the authors.

Licence

Released under the Apache License 2.0, the same licence as the original (see LICENSE). Edda is a fine-tune of openai/whisper-large-v3-turbo, which is MIT-licensed (see LICENSE-whisper). Attribution and the modification notice are in NOTICE.

What was changed

The weights were converted, not retrained:

1. fp16 safetensors to GGML f16 with CrispASR v0.8.29's models/convert-h5-to-ggml.py. The source repo ships only tokenizer.json, so vocab.json and added_tokens.json were derived from it for the converter; both are identical to openai/whisper-large-v3-turbo's. Because added_tokens.json is present, the file's vocabulary carries the 1,609 special-token strings (<|da|>, <|0.00|>, ...) that whisper.cpp's own conversions leave for the runtime to generate. That is the only difference from a stock conversion: 20,164 bytes of header, tensors unchanged.
2. f16 to q8_0 and q5_0 with CrispASR's legacy Whisper quantizer (crispasr-legacy-quantize).

Files

File	Notes
ggml-edda-v0.2-f16.bin	full precision, largest and slowest
ggml-edda-v0.2-q8_0.bin	about half the size of f16
ggml-edda-v0.2-q5_0.bin	smallest and fastest

Codictate's own measurements on FLEURS da_dk are in its repository. ggml-edda-v0.2-q5_0.bin is the place to start.

Runtime

These are legacy GGML Whisper files. They load in whisper.cpp and in CrispASR's default Whisper backend. Pass --language da: the model is Danish only.

Requests

If the Alexandra Institute or Danish Foundation Models would like the attribution worded differently, or would prefer this mirror taken down, open an issue on the Codictate repository and it will be removed.