metadata
license: mit
tags:
- phi3
- nlp
- moe
datasets:
- BEE-spoke-data/gutenberg-en-v1-clean
- NeelNanda/pile-10k
phi 3 4x4b
a continually pretrained phi3-mini sparse moe upcycle
benchmarks
ran locally
Microsoft/phi-3-4k-instruct | Fizzarolli/phi3-4x4b-v1 | |
---|---|---|
MMLU acc. (0-shot) | 0.6799 | 0.6781 |
Hellaswag acc. (0-shot) | 0.6053 | 0.5962 |
ARC-E acc. (0-shot) | 0.8325 | 0.8367 |
ARC-C acc. (0-shot) | 0.5546 | 0.5606 |
honestly i was expecting it to do worse :p, but those are all within a margin of error! so it didn't lose any performance, at least
open llm leaderboard
todo!
support me on ko-fi!
please i need money to stay alive and keep making models
notes
not trained on instruct data. it's pretty likely that it won't be much different from phi 3 if you use it like that, if not worse due to any forgetting of instruct formats during the continued training.
future experiments
- the datasets for this were literally chosen on a whim. perhaps experiment with a further filtered HuggingFaceFW/fineweb-edu?
- actually freeze the gate layers next time (see Chen et. al, 2023), oops
- MOAR TRAINING, this only went up to ~0.2 of an epoch because i ran out of dolar