Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
SingL3 709bb99d3e | 8 months ago | |
---|---|---|
.. | ||
model_eval | 8 months ago | |
model_training | 8 months ago | |
pretokenizer | 9 months ago | |
.gitignore | 1 year ago | |
MESSAGE_AND_TOKEN_FORMAT.md | 11 months ago | |
README.md | 10 months ago | |
pyproject.toml | 9 months ago |
Here are some minimal commands to tun to whole pipeline on the collected data.
make sure python >= 3.10, otherwise, you would meet the
[issue]
mkdir -p .cache
mkdir -p .saved_models
export DATA_PATH=$PWD/.cache
export MODEL_PATH=$PWD/.saved_models
Create a new or modify an existing configuration section in the config.yaml
(SFT), config_rm.yaml
(RM) or config_rl.yaml
(RL) YAML configuration files
located in the model_training/configs/
directory and specify the OA JSONL data
file or HuggingFace dataset to use.
.jsonl
or .jsonl.gz
) specify theinput_file_path
configuration option. Place the filecache_dir
(DATA_PATH
) or specify an absolute path.cp /path/to/<oasst.trees.jsonl> $DATA_PATH
Example:
my_data_config:
datasets:
- oasst_export:
input_file_path: oasst_export.trees.jsonl.gz
hf_dataset_name
configuration option.Example:
my_data_config:
datasets:
- oasst_export:
hf_dataset_name: OpenAssistant/oasst1
Note: If both hf_dataset_name
and input_file_path
are specified
input_file_path
will take precedence.
See the
OpenAssistant/oasst1
dataset card on the HuggingFace hub for more information.
cd model_training
# export shared modules
export PYTHONPATH=$PYTHONPATH:../../oasst-shared
python trainer_sft.py --configs defaults oa_dataset_only pythia --cache_dir $DATA_PATH --output_dir $MODEL_PATH/sft_model
# if you want to use wandb, add
--wandb_entity your_username/team_name
To change the model used, i.e. larger pythia version create a new config in
model_training/configs/config.yaml
or set the flag --model_name
to
EleutherAI/pythia-{size}-deduped
. Larger models will probably need to also
adjust the --learning_rate
and --per_device_train_batch_size
flags.
# choose a specific checkpoint
export SFT_MODEL=$MODEL_PATH/sft_model/<checkpoint-X>
# or get latest checkpoint
export SFT_MODEL=$MODEL_PATH/sft_model/$(ls -t $MODEL_PATH/sft_model/ | head -n 1)
cd model_training
python trainer_rm.py --configs defaults_rm oasst-rm-1-pythia-1b
# choose a specific checkpoint
export REWARD_MODEL=$MODEL_PATH/reward_model/<checkpoint-X>
# or get latest checkpoint
export REWARD_MODEL=$MODEL_PATH/reward_model/$(ls -t $MODEL_PATH/reward_model/ | head -n 1)
cd model_training
python trainer_rl.py --configs defaults_rlhf --cache_dir $DATA_PATH --rank_model $REWARD_MODEL --sft_model $SFT_MODEL --output_dir $MODEL_PATH/rl_model
See the MESSAGE_AND_TOKEN_FORMAT.md
file for information about the pattern we
are using.
No Description
Jupyter Notebook Python TSX TypeScript SVG other
Dear OpenI User
Thank you for your continuous support to the Openl Qizhi Community AI Collaboration Platform. In order to protect your usage rights and ensure network security, we updated the Openl Qizhi Community AI Collaboration Platform Usage Agreement in January 2024. The updated agreement specifies that users are prohibited from using intranet penetration tools. After you click "Agree and continue", you can continue to use our services. Thank you for your cooperation and understanding.
For more agreement content, please refer to the《Openl Qizhi Community AI Collaboration Platform Usage Agreement》