Download notes/progress.md from hulehule/pllm2-full-dump: direct link, hf CLI and curl.
- Browser
- Download file 2.76 kB
-
https://huggingface.co/hulehule/pllm2-full-dump/resolve/main/notes/progress.md
- Command line
-
hf download hf://hulehule/pllm2-full-dump/notes/progress.md
-
curl -L -o progress.md https://huggingface.co/hulehule/pllm2-full-dump/resolve/main/notes/progress.md
July 8:
- link to HDFS
- HDFS address: /mnt/hdfs/pllm/yutong
- uploaded two benchmarks
- downloaded the long_lamp datasets
HDFS删除文件: hadoop fs -rm -r /hdfs/文件夹/路径 删除单个文件: hadoop fs -rm /hdfs/文件/路径
baseline training for abtract genration:
PYTHONPATH=. python3 train_llm.py
--train_data /mnt/hdfs/pllm/yutong/long_lamp/abstract_train.json
--validation_data /mnt/hdfs/pllm/yutong/long_lamp/abstract_train.json
--test_data /mnt/hdfs/pllm/yutong/long_lamp/abstract_test.json
--model_name google/flan-t5-base
--task generation_abstract
--output_dir /mnt/hdfs/pllm/yutong/long_output
#多卡训练
torchrun --nproc_per_node=4 --master_port=29502 train_llm.py
--train_data /mnt/hdfs/pllm/yutong/long_lamp/abstract_train.json
--validation_data /mnt/hdfs/pllm/yutong/long_lamp/abstract_train.json
--test_data /mnt/hdfs/pllm/yutong/long_lamp/abstract_test.json
--model_name google/flan-t5-base
--task generation_abstract
--output_dir /mnt/hdfs/pllm/yutong/long_output
#改成llama 3.1 8B 版本
PYTHONPATH=. python3 train_llama.py
--train_data /mnt/hdfs/pllm/yutong/long_lamp/abstract_train.json
--validation_data /mnt/hdfs/pllm/yutong/long_lamp/abstract_train.json
--test_data /mnt/hdfs/pllm/yutong/long_lamp/abstract_test.json
--model_name meta-llama/Meta-Llama-3.1-8B
--task generation_abstract
--output_dir /mnt/hdfs/pllm/yutong/long_output_llama
HDFS 删除文件: hdfs dfs -rm -r harunava/home/byte_tiktok_content_ecology/user/yutong /mnt/hdfs/pllm/abstract_train.json hdfs dfs -rm /mnt/hdfs/pllm/abstract_test.json
HDFS 删除文件: hdfs dfs -rm -r harunava/home/byte_tiktok_content_ecology/user/yutong /mnt/hdfs/pllm/abstract_train.json hdfs dfs -rm /mnt/hdfs/pllm/abstract_test.json
July 16:
- baseline model using Llama 3.1
- model 2: RL for retrieval refinement
July 17:
- 重新下载lamp数据集
- run long- writer在long_lamp数据集上