search

mistralai / mistral-inference

Official inference library for Mistral models
https://mistral.ai/
Apache License 2.0
9.16k stars 803 forks source link

Training code #138

Open sartimo opened 3 months ago

sartimo commented 3 months ago

Hi

Where can I find the code needed to train the initial model and produce the model files?

spydaz commented 3 months ago

Hi

Where can I find the code needed to train the initial model and produce the model files?

Yes to instanciate a fresh model from te code and train a new model (after generating the initial config(ie Load with no weights).... as i would like to train a model from scratch with my own data (i have a trained tokenizer but i would also like a script to train a new tokenizer)

This is needed to create a new model ( with a new tokenization process, ie : multimodal input ) ... so the ability to select which input pre processors / feature extractors are available ... As Speech input should be auto tokenized ... from transcribed to text to token_ID... as well as the image being returned to token _IDs also , for images the processor would process the image and convert to token ID... In training the image may have relevance as reference as it should give a description of an image to the prompt( a man sitting on a bench), (a Malignant tumour with legions) .. so again ot would be tokenized to words.... the words need to transcribe and the images need to return their learned captions. when traning the peft would be applied to the transformer ! ... the image peft applied to the processor and the sound applied to the open whisper processor etc...

We need only to have a text output , as later we can create a wrapper for generation of sound and for images ... using the same sound but with diffusers to generate an output. the training process should use the diffusers to learn the images as well as by the captioning its description .... hence for later generation any pre captioned image should be able to be regenerated or a representation ! ... For sound input and generation , obviously speech output is no problem as the same library for speech also outputs speech, but we also need a sound generator for our generated outputs ie generate the sound of a sparrow(bird).... (another form of diffuser)....

hence we need the start location ! the training script for a code model ... ?