mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-11-03 02:13:07 +00:00
74069db498
* Reorganize the data preparation scripts for tulu v1 and v2. * Minor improvement * Remove open_platypus_commercial subset from Daring-Anteater * Use hard-coded examples repo. * Fix some bugs. * Add OpenMathInstruct. * Add a few more v3.5.x SFT mix ablations for the cleaner datasets. * More experiments on mixes. * help merge * prep for merge * reapply changes * fix naming --------- Co-authored-by: Nathan Lambert <[email protected]>
32 lines
1.3 KiB
YAML
32 lines
1.3 KiB
YAML
model_name_or_path: meta-llama/Llama-3.1-8B
|
|
model_revision: main
|
|
use_flash_attn: true
|
|
tokenizer_name: meta-llama/Llama-3.1-8B
|
|
use_slow_tokenizer: true
|
|
dataset_mixer:
|
|
# Remove all mis-licensed data or non commercial
|
|
ai2-adapt-dev/metamath-qa-reformat: 1.0 # MIT License
|
|
natolambert/tulu-v2-sft-mixture-flan: 1.0 # FLAN Apache 2.0
|
|
natolambert/tulu-v2-sft-mixture-cot: 1.0 # FLAN Apache 2.0
|
|
allenai/openassistant-guanaco-reformatted: 1.0 # Apache 2.0
|
|
ai2-adapt-dev/codefeedback-single-turn-reformat-magicoder: 1.0 # MIT MagiCoder section of CodeFeedback
|
|
ai2-adapt-dev/aya_dataset-reformat: 1.0 # Apache 2.0
|
|
ai2-adapt-dev/SlimOrca-reformat: 0.25 # MIT License
|
|
ai2-adapt-dev/Daring-Anteater-reformat: 1.0 # CC BY 4.0
|
|
ai2-adapt-dev/WebInstructSub-reformat-apache: 0.1 # Apache 2.0
|
|
ai2-adapt-dev/Table-GPT-All-train: 0.5 # MITmax_seq_length: 4096 # Note, reduced from 8192 to fit on one GPU with DeepSpeed Stage3
|
|
preprocessing_num_workers: 128
|
|
per_device_train_batch_size: 1 # note, this is set up for 8 GPUs
|
|
gradient_accumulation_steps: 2 # effective batch size 128 with 1 node
|
|
learning_rate: 5.0e-06 # best LR so far
|
|
lr_scheduler_type: linear
|
|
warmup_ratio: 0.03
|
|
weight_decay: 0.0
|
|
num_train_epochs: 2
|
|
output_dir: /output/
|
|
with_tracking: true
|
|
report_to:
|
|
- wandb
|
|
logging_steps: 1
|
|
checkpointing_steps: epoch
|
|
dataset_mix_dir: /output/ |