ADBL2-Mistral-7B

ADBL2-Mistral-7B is an fine-tuned version of Mistral-7B-v0.1 trained to perform relation-based argument mining. Giving two arguments x and y, we use this model in synergy with LMQL to predict wether y is attacking or supporting x.

Fine-tuning

We fine-tunde Mistral-7B-v0.1 using the PEFT method QLoRA and argument pairs coming from the online debate tool Kialo.

Prompt format

This model has been trained to complete this prompt format:

<s>[INST]
Argument 1 : /*Argument 1*/
Argument 2 : /*Argument 2*/
[/INST]
Relation : 

by the relation attack or support

Relation : attack/support
</s>

Example :

Giving two arguments, where argument 2 is attacking the argument 1, :

• Argument 1 : using machines is advantageous
• Argument 2 : the usage of machines is harmful for health of humans

The prompt to retrieve the relation between the second and the first argument should be :

<s>[INST]
Argument 1 : using machines is advantageous
Argument 2 : the usage of machines is harmful for health of humans
[/INST]
Relation : 

Our model should complete this prompt this way :

<s>[INST]
Argument 1 : using machines is advantageous
Argument 2 : the usage of machines is harmful for health of humans
[/INST]
Relation : attack
</s>