mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-11-03 02:13:07 +00:00
57 lines
2.0 KiB
YAML
57 lines
2.0 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:
|
|
# Static v3.9 huggingface dataset
|
|
allenai/tulu-v.3.9-mix-preview-noncommercial: 1.0
|
|
|
|
# # General datasets:
|
|
# ai2-adapt-dev/oasst1_converted: 1.0 # 7132 # all
|
|
# ai2-adapt-dev/flan_v2_converted: 1.0 # 89982 # all
|
|
# ai2-adapt-dev/tulu_hard_coded_repeated_10: 1.0 # 240 # all
|
|
# ai2-adapt-dev/no_robots_converted: 1.0 # 9500 # all
|
|
# ai2-adapt-dev/tulu_v3.9_wildchat_100k: 1.0
|
|
|
|
# # Math datasets:
|
|
# ai2-adapt-dev/personahub_math_v5_regen_149960: 1.0 # 149960 # all
|
|
# allenai/tulu-3-sft-personas-math-grade: 1.0 # 49980 # all
|
|
# ai2-adapt-dev/tulu_v3.9_open_math_2_gsm8k_50k: 1.0
|
|
# ai2-adapt-dev/numinamath_tir_math_decontaminated: 1.0
|
|
# ai2-adapt-dev/tulu_v3.9_personahub_math_interm_algebra_20k: 1.0
|
|
|
|
# # Coding datasets:
|
|
# ai2-adapt-dev/personahub_code_v2_34999: 1.0 # 34999 # all
|
|
# ai2-adapt-dev/evol_codealpaca_heval_decontaminated: 1.0 # 107276 # all
|
|
|
|
# # IF datasets:
|
|
# ai2-adapt-dev/personahub_ifdata_manual_seed_v3_29980: 1.0 # 29980 # all
|
|
|
|
# # Safety datasets:
|
|
# ai2-adapt-dev/coconot_converted: 1.0 # 10983 # all
|
|
# ai2-adapt-dev/tulu_v3.9_wildjailbreak_decontaminated_50k: 1.0
|
|
# ai2-adapt-dev/tulu_v3.9_synthetic_finalresp_wildguardmixtrain_decontaminated_50k: 1.0
|
|
|
|
# # Specialty datasets:
|
|
# ai2-adapt-dev/tulu_v3.9_sciriff_10k: 1.0
|
|
# ai2-adapt-dev/tulu_v3.9_table_gpt_5k: 1.0
|
|
# ai2-adapt-dev/tulu_v3.9_aya_100k: 1.0
|
|
|
|
max_seq_length: 4096
|
|
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 8 nodes
|
|
learning_rate: 5.0e-06
|
|
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: 1000
|
|
keep_last_n_checkpoints: 1
|
|
dataset_mix_dir: /output/ |