KamilHugsFaces / t5-gemma-2-reasoning-classifier-v3

KamilHugsFaces/t5-gemma-2-reasoning-classifier-v3

Fine-tuned T5-Gemma-2 model for entailment classification.

Training Details

• Base model : google/t5gemma-2-4b-4b
• Training variant : reasoning_v1
• Epochs : 3
• Batch size : 4
• Learning rate : 5e-05
• Run name : reasoning_v1_20260113_222305

Training Data

• Training examples : 700
• Validation examples : 150
• Test examples : 150
• Class weights : {'true': 0.5417118093174431, 'false': 6.4935064935064934}

Evaluation Results

Test Set Performance

• F1 Score : 0.8120
• F1 (False class) : 0.3333
• Accuracy : 0.7600
• Precision (False) : 0.2143
• Recall (False) : 0.7500

Usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("KamilHugsFaces/t5-gemma-2-reasoning-classifier-v3")
tokenizer = AutoTokenizer.from_pretrained("KamilHugsFaces/t5-gemma-2-reasoning-classifier-v3")

# Format input
input_text = "entailment: [Your claim and evidence here]"
inputs = tokenizer(input_text, return_tensors="pt", max_length=250, truncation=True)

# Generate prediction
outputs = model.generate(**inputs, max_new_tokens=8)
prediction = tokenizer.decode(outputs[0], skip_special_tokens=True)
# Output: "true" or "false"

Training Configuration

{ "variant_name": "reasoning_v1", "run_name": "reasoning_v1_20260113_222305", "num_epochs": 3, "batch_size": 4, "learning_rate": 5e-05, "warmup_steps": 100, "model_name": "google/t5gemma-2-4b-4b", "class_weights": { "true": 0.5417118093174431, "false": 6.4935064935064934 }, "use_confidence_weighting": false, "confidence_weight_alpha": 2, "train_size": 700, "val_size": 150, "test_size": 150 }

Framework

• Transformers : 5.0.0.dev0
• PyTorch : 2.9.1+cu128
• Trained on : Modal (A100 GPU)