cognitivecomputations

/

dolphin-2.1-mistral-7b

like 255

It's nice seeing all the developments chug along, with no stopping in sight! Thank you for your continued work in training these models. Being able to see and experience what's on the frontier for LLMs first-hand is really valuable. I had a few questions about the training details for this (very coherent) model, with potential areas for improvement:

As per the original Orca paper, the training for their model used GPT-3.5 completions as an intermediary target to meet before later training on GPT-4 completions, making sure the model has been finalized with the higher quality data (from GPT-4) at its training completion. Relevant excerpt from paper:

We first train Orca on FLAN-5M (ChatGPT augmentations), followed by second stage of training on FLAN-1M (GPT-4 augmentations). Essentially, we leverage ChatGPT as intermediate teacher assistant for two reasons.

• Capacity gap: Orca with 13B parameters is many times smaller than GPT-4 (size undisclosed). Leveraging an intermediate teacher with reduced gap in capabilities, in this case ChatGPT, has been shown to improve imitation learning performance for smaller students in knowledge distillation [15]. This can be viewed as a form of progressive learning or curriculum learning, where the student first learns from easier examples, followed by harder ones: with the assumption that longer responses are difficult to mimic than shorter ones, along with improved reasoning and step-by-step explanation from a larger teacher . . .

From what I could gather, the training for dolphin-2.1 was performed with a mix of data in no particular order, which would suggest that the model learns using that mixed-quality target dataset all the way to training completion. If this is the case, future training could benefit from a more ordered learning process, for instance: 3 epochs of GPT-3.5 completions (w. shuffling on each epoch and a suitable LR scheduler) -> 5 epochs of GPT-4 completions, where airoboros data (which is from GPT-4) could be mixed into.

Separation of training into these stages, with each stage having its own scheduling — both leading to a final LR cool-down — can also act in a way similar to cosine annealing, refreshing the training and providing a chance for escaping local minima (thereby achieving better generalization). This would also add more hyperparameters to fiddle with, both the number of epochs and peak LR for each stage now variable.

The other question is on the filtering for both GPT completions: was the deduplication step in filtering performed per-set or across both GPT-4 and GPT-3.5 sets? Deduplication should have been performed per-set with GPT-3.5 and GPT-4 completions treated as completely separate datasets, given the curriculum learning intentions.

Last is a small question about the formatting. It seems the model outputs for the assistant role always begin with a token containing a space character (logged from llama.cpp, notice token 6880):

'':32000, '':32001, 'ass':489, 'istant':11143, '':13, ' Another':6880, ' lesser':26767, '-':28733, 'known':4717, ' but':562, ' fascinating':23069, ' quote':13658, ' from':477, ' Charles':6427

Was this an intentional decision (as with default Llama tokenizer behavior, adding BOS and having the very first token include a space), and should the user input also include a space?