[TEST-USE] Fine Tuning

[guerra2fernando]

[jieguangzhou]

Environment (with MongoDB + 1 type of data + 1 model):

• [x] local cluster
• [x] development mode

DBs (development mode + 1 non-trivial type of data + 1 model):

• [x] MongoDB
• [x] SQLite
• [x] Postgres

Data Format (MongoDB + 1 model + development mode)

• [x] Chat
• [x] Prompt-Response
• [x] Prompt

Training options (MongoDB + 1 type of data + development mode)

• [x] ray
• [x] deepspeed
• [x] lora
• [x] qlora
• [x] Multi-GPUS

Load Options

• [x] Directly
• [x] Use checkpoint

[jieguangzhou]

At the functional level, the test has been completed (running a small max_steps test). Because it involves model training, work at the detection effect level is required. I think I need to continue to run the real finetune task on data in different formats to test the effect after the model training is completed, Each training takes a long time, but other work can be done simultaneously

Data Format (MongoDB + 1 model + development mode)

• [x] Chat
• [x] Prompt-Response
• [x] Prompt