Foundation Model

[lss11005]

Can I use other model to replace [meta-llama/Llama-2-7b-hf], such as Mistral-7b?
which part should to be modified if I can replace, data preprocess or other config~~

[Spico197]

Sure, you can use other models by setting --model_name_or_path in train_seed.sh and train_backward_Myx.sh.

[lss11005]

Should other training parameters need to be modified? such as learning_rate, beta, training data~~
Have you tried other models?
I finetuned model(zephyr-7b-sft-full) with seed.jsonl, but test acc drop a lot( test on Winogrande )
zephyr-7b-sft-full 0.7419
m0 0.5667

[Spico197]

No I didn't try on other models. Since you are finetuning on an SFT model, you may want to change the conversation template accordingly. e.g. https://github.com/Spico197/Humback/blob/641153764e59bd4443a2c42b893c9fec6db24467/src/data.py#L17

[lss11005]

template is necessary? can I construct data myself, such as "instruction" -- "response“； and reverse while train M_yx

[Spico197]

Sure you can construct dataset yourself. Just need to change the corresponding data construction process accordingly.