How to do training data preprocessing and templates (non-conversational data training)?

[supperman009]

I want Qwen2.5 or DeepSeek R1 4Bit quantitative model to learn the basic table structure, table name, field name, and field explanation of more than 200 tables in the database of my business system through fine-tuning training. Then use the basic capabilities of the model to analyze and answer questions about these tables.
However, no matter how I organize the training data, use instructions, input and output dialogue templates, and fine-tune Lora, the model either cannot learn information related to the table, or the model is overfitted and can only answer questions related to the table.
I think it may be that my data preprocessing is wrong? Or the training data is wrong? Or the training template is wrong?
I checked some information, and some people suggested that the basic model should be trained in unsupervised text without organizing it into a dialogue instruction data set. Can Unsloth perform such fine-tuning training on the model? Can you provide an example using Qwen2.5 as an example?
Unsloth's current examples are mainly instructions to fine-tune SFTTrainer. I found some papers, which mean that such SFT fine-tuning belongs to supervised fine-tuning. The training data is mainly high-quality and diverse dialogue data. It is through supervised learning that the model can better answer questions based on the knowledge content that the basic model has trained. Including the reasoning model of DeepSeek R1, they are all similar.
So I am very confused. I want the model to learn the information of more than 200 tables so that in the reasoning process, the model can generalize its reasoning ability to these table data. However, the SFT training mode does not seem to allow the model to learn the knowledge information of the table. It can only allow the model to learn the template in the form of question and answer.

Who has tried this aspect? Can you provide training examples and how to fine-tune this type of model scenario?
Thank you!

[alberto071968]

Very interesting topic. I would like to tackle this problem as well. Thanks Il ven 28 feb 2025, 16:00 supperman009 ***@***.***> ha scritto:

…

I want Qwen2.5 or DeepSeek R1 4Bit quantitative model to learn the basic table structure, table name, field name, and field explanation of more than 200 tables in the database of my business system through fine-tuning training. Then use the basic capabilities of the model to analyze and answer questions about these tables. However, no matter how I organize the training data, use instructions, input and output dialogue templates, and fine-tune Lora, the model either cannot learn information related to the table, or the model is overfitted and can only answer questions related to the table. I think it may be that my data preprocessing is wrong? Or the training data is wrong? Or the training template is wrong? I checked some information, and some people suggested that the basic model should be trained in unsupervised text without organizing it into a dialogue instruction data set. Can Unsloth perform such fine-tuning training on the model? Can you provide an example using Qwen2.5 as an example? Unsloth's current examples are mainly instructions to fine-tune SFTTrainer. I found some papers, which mean that such SFT fine-tuning belongs to supervised fine-tuning. The training data is mainly high-quality and diverse dialogue data. It is through supervised learning that the model can better answer questions based on the knowledge content that the basic model has trained. Including the reasoning model of DeepSeek R1, they are all similar. So I am very confused. I want the model to learn the information of more than 200 tables so that in the reasoning process, the model can generalize its reasoning ability to these table data. However, the SFT training mode does not seem to allow the model to learn the knowledge information of the table. It can only allow the model to learn the template in the form of question and answer. Who has tried this aspect? Can you provide training examples and how to fine-tune this type of model scenario? Thank you! — Reply to this email directly, view it on GitHub <#1857>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BFBQP7JDXRCPNAJ5KFGFXRT2SB2XBAVCNFSM6AAAAABYCQGSTKVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZYGAZDKMRUGA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>