finetune_name = 'Askinkaty/llama-finance-relations'


finetined_model = AutoPeftModelForCausalLM.from_pretrained(
    pretrained_model_name_or_path=finetune_name,
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True,
    device_map="auto",
)

base_model_name = "meta-llama/Llama-3.2-1B-Instruct"
base_model = AutoModelForCausalLM.from_pretrained(base_model_name,
                                                  torch_dtype=torch.float16,
                                                  low_cpu_mem_usage=True)

tokenizer = AutoTokenizer.from_pretrained(model_name)

base_model.config.pad_token_id = base_model.config.eos_token_id
model.config.pad_token_id = model.config.eos_token_id

pipeline = pipeline('text-generation', model=base_model, tokenizer=tokenizer, max_length=1024, device=device)
pipeline.model = model.to(device) 

Training Details

Training Data

Samples from ReFinD dataset. 100 examples for each relation type were used, least frequent relation types are omitted.

Preprocessing

Dataset is converted into a message format as in the code snippet below:

python

def batch_convert_to_messages(data):
        
    questions = data.apply(
        lambda x: f"Entity 1: {' '.join(x['token'][x['e1_start']:x['e1_end']])}. "
                  f"Entity 2: {' '.join(x['token'][x['e2_start']:x['e2_end']])}. "
                  f"Input sentence: {' '.join(x['token'])}",
        axis=1
    )

    relations = data['relation'].apply(lambda relation: relation.split(':')[-1])
    
    messages = [
        [
            {
                "role": "system",
                "content": "You are an expert in financial documentation and market analysis. Define relations between two specified entities: entity 1 [E1] and entity 2 [E2] in a sentence. Return a short response of the required format. "
            },
            {"role": "user", "content": question},
            {"role": "assistant", "content": relation},
        ]
        for question, relation in zip(questions, relations)
    ]
    
    return messages

Training Hyperparameters

SFT parameters:

num_train_epochs=1
per_device_train_batch_size=2
gradient_accumulation_steps=2
gradient_checkpointing=True
optim="adamw_torch_fused"
learning_rate=2e-4
max_grad_norm=0.3
warmup_ratio=0.01
lr_scheduler_type="cosine"
bf16=True

LORA parameters:

rank_dimension = 6
lora_alpha = 8
lora_dropout = 0.05

Evaluation

Testing Data

Test set sampled from Samples from ReFinD dataset.

Metrics

Overall Performance:
Precision: 0.77
Recall: 0.69
F1 Score: 0.71

Classification Report:
                  precision    recall  f1-score   support

     no_relation       0.00      0.00      0.00         0
           title       0.00      0.00      0.00         0
   operations_in       0.65      0.66      0.66       100
     employee_of       0.00      0.00      0.00         0
  agreement_with       0.58      0.88      0.70       100
       formed_on       0.00      0.00      0.00         0
       member_of       0.99      0.96      0.97        96
   subsidiary_of       0.00      0.00      0.00         0
       shares_of       0.00      0.00      0.00         0
      revenue_of       0.60      0.27      0.38        95
         loss_of       0.64      0.37      0.47       100
headquartered_in       0.99      0.73      0.84       100
     acquired_on       0.00      0.00      0.00         0
      founder_of       0.74      0.77      0.76        83
       formed_in       0.96      0.91      0.93       100

        accuracy                           0.69       774
       macro avg       0.41      0.37      0.38       774
    weighted avg       0.77      0.69      0.71       774

PEFT 0.14.0