metadata
license: mit
pipeline_tag: text-generation
library_name: transformers
language:
- en
- am
- ar
- as
- az
- be
- bg
- bn
- br
- bs
- ca
- cs
- cy
- da
- de
- el
- eo
- es
- et
- eu
- fa
- ff
- fi
- fr
- fy
- ga
- gd
- gl
- gn
- gu
- ha
- he
- hi
- hr
- ht
- hu
- hy
- id
- ig
- is
- it
- ja
- jv
- ka
- kk
- km
- kn
- ko
- ku
- ky
- la
- lg
- li
- ln
- lo
- lt
- lv
- mg
- mk
- ml
- mn
- mr
- ms
- my
- ne
- nl
- 'no'
- ns
- om
- or
- pa
- pl
- ps
- pt
- qu
- rm
- ro
- ru
- sa
- si
- sc
- sd
- sk
- sl
- so
- sq
- sr
- ss
- su
- sv
- sw
- ta
- te
- th
- tl
- tn
- tr
- ug
- uk
- ur
- uz
- vi
- wo
- xh
- yi
- yo
- zu
datasets:
- ontocord/fineweb-permissive-multilingual-2m
- distily/c4_multilingual_1M
- data-silence/sumnews
- xu-song/cc100-samples
- badrex/llm-emoji-dataset
- fblgit/simple-math
- Gusarich/math-expressions-1m
- neuralwork/arxiver
- christopher/rosetta-code
- nampdn-ai/tiny-codes
- JeanKaddour/minipile
- NousResearch/hermes-function-calling-v1
- simplescaling/s1K-1.1
- mlabonne/open-perfectblend
- allenai/tulu-3-sft-mixture
- rombodawg/Everything_Instruct_Multilingual
- open-r1/OpenR1-Math-220k
- open-thoughts/OpenThoughts-114k
- cognitivecomputations/dolphin-r1
- simplescaling/s1K-1.1
tags:
- chat
- core
- base
- instruct
- reason
tangled-alpha-0.12-core
time python -B prepare_base_datasets.py
i=0, min_len=0, max_len=1073741824, block_size=8193, chunk_size=16386000, len(dataset)=1496631, len(dataset) * block_size=12261897783
Total number of tokens in the optimized dataset '../base-data-0-0-1073741824-8193-2000' is 12261897783
i=1, min_len=8193, max_len=16385, block_size=16385, chunk_size=16385000, len(dataset)=78802, len(dataset) * block_size=1291170770
Total number of tokens in the optimized dataset '../base-data-1-8193-16385-16385-1000' is 1291170770
i=2, min_len=16385, max_len=32769, block_size=32769, chunk_size=16384500, len(dataset)=23511, len(dataset) * block_size=770431959
Total number of tokens in the optimized dataset '../base-data-2-16385-32769-32769-500' is 770431959
i=3, min_len=32769, max_len=65537, block_size=65537, chunk_size=16384250, len(dataset)=5128, len(dataset) * block_size=336073736
Total number of tokens in the optimized dataset '../base-data-3-32769-65537-65537-250' is 336073736
i=4, min_len=65537, max_len=131073, block_size=131073, chunk_size=16384125, len(dataset)=1169, len(dataset) * block_size=153224337
Total number of tokens in the optimized dataset '../base-data-4-65537-131073-131073-125' is 153224337
46G ../base-data-0-0-1073741824-8193-2000
4.9G ../base-data-1-8193-16385-16385-1000
2.9G ../base-data-2-16385-32769-32769-500
1.3G ../base-data-3-32769-65537-65537-250
589M ../base-data-4-65537-131073-131073-125
CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=0 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True litgpt pretrain --config pretrain_base_model_0.yaml
Backup wandb
:
mv wandb wandb-pretrain-base-0
Copy config:
cp ../config-0.json ../out/pretrain-base-0/final/config.json
Chat with model:
CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=0 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True litgpt chat ../out/pretrain-base-0/final
CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=0 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True time litgpt evaluate --tasks 'leaderboard' --out_dir '../evaluate/pretrain-base-0/leaderboard/' --batch_size '4' --dtype 'bfloat16' '../out/pretrain-base-0/final'
litgpt convert_pretrained_checkpoint ../out/pretrain-base-0/final ../out/pretrain-base-0/checkpoint
CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=0 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True litgpt pretrain --config pretrain_base_model_1.yaml
litgpt convert_pretrained_checkpoint ../out/pretrain-base-1/final ../out/pretrain-base-1/checkpoint
CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=0 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True litgpt pretrain --config pretrain_base_model_2.yaml
litgpt convert_pretrained_checkpoint ../out/pretrain-base-2/final ../out/pretrain-base-2/checkpoint
CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=0 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True litgpt pretrain --config pretrain_base_model_3.yaml
CUDA_VISIBLE_DEVICES=0 CUDA_LAUNCH_BLOCKING=0 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True time litgpt evaluate --tasks 'leaderboard' --out_dir '../evaluate/pretrain-base-3/leaderboard/' --batch_size '4' --dtype 'bfloat16' '../out/pretrain-base-3/final'