SFT for D2L + Pre-Training (rename of the previous SFT)#102

Open

llauraa23 wants to merge 16 commits into

CambioML:mainfrom

llauraa23:sft-d2l

Commits on Dec 5, 2023

auto rater, sample data and prompt engineering
llauraa23
committed

Commits on Dec 30, 2023

Merge branch 'CambioML:main' into main
llauraa23
authored

Commits on Jan 29, 2024

merge conflict

llauraa23
authored and
Diyi Hu
committed
support supervised fine tuning on d2l.

llauraa23
authored and
Diyi Hu
committed
resolve merge conflicts on gpu96

llauraa23
authored and
Diyi Hu
committed
support training multiple epochs in sft.

llauraa23
authored and
Diyi Hu
committed
implment evaluation of fine-tuned models with pykoi pipeline

llauraa23
authored and
Diyi Hu
committed
When evaluating the SFT model, store the questions and generated answers into a csv file

llauraa23
authored and
Diyi Hu
committed
code cleanup for d2l demo. In SFT, make data collator, formatting function, whether to disable evalution configurable

llauraa23
authored and
Diyi Hu
committed
rename previous sft as pre-training. Modify corresponding demo scripts/notebooks
Diyi Hu
committed
In sft, make the number of entries in data for training configurable. Use trl DataCollatorForCompletionOnlyLM instead of customized one. Debug: cannot use ConstantLengthDataset or packing when usin…
Diyi Hu
committed

Commits on Feb 6, 2024

remove redundant implementation of autorater
llauraa23
committed

Commits on Feb 7, 2024

Commits on Feb 14, 2024