| # Model Card: t5-summary-finetuned-kw-fp16 | |
| # Model Overview | |
| - **Model Name:** t5-summary-finetuned-kw-fp16 | |
| - **Base Model:** T5-base (t5-base from Hugging Face) | |
| - **Date:** March 19, 2025 | |
| - **Version:** 1.0 | |
| - **Task:** Keyword-Based Text Summarization | |
| - **Description:** A fine-tuned T5-base model quantized to FP16 for generating concise summaries from short text inputs, guided by a user-specified keyword. Trained on a custom dataset of 200 examples, it produces summaries focusing on the keyword while maintaining a professional tone. | |
| # Model Details | |
| - **Architecture:** Encoder-Decoder Transformer (T5-base) | |
| - **Parameters:** ~223M (original T5-base), quantized to FP16 | |
| - **Precision:** FP16 (16-bit floating-point) | |
| - **Input Format:** Text paragraph + "Keyword: [keyword]" (e.g., "The storm caused heavy rain and wind damage. Keyword: rain") | |
| - **Output Format:** Concise summary (1-2 sentences) focusing on the keyword (e.g., "The storm brought heavy rain overnight.") | |
| - **Training Hardware:** NVIDIA GPU with 12 GB VRAM (e.g., RTX 3060) | |
| - **Inference Hardware:** Compatible with GPUs supporting FP16 (minimum ~1.5 GB VRAM) | |
| # Training Data | |
| **Dataset Name:** Custom Keyword-Based Summarization Dataset | |
| - **Size:** 200 examples | |
| - **Split:** 180 training, 20 validation | |
| - **Format:** CSV | |
| - **input:** Paragraph (2-4 sentences) + "Keyword: [keyword]" | |
| - **keyword:** Single word or short phrase guiding the summary | |
| - **output:** Target summary (1-2 sentences) | |
| - **Content:** Diverse topics including tech, weather, sports, health, and culture (e.g., "A new laptop was released with a fast processor... Keyword: processor" → "The new laptop has a fast processor.") | |
| - **Language:** English | |
| # Training Procedure | |
| - **Framework:** PyTorch via Hugging Face Transformers | |
| # Hyperparameters: | |
| **Epochs:** 2 (stopped early; originally set for 3) | |
| - **Learning Rate:** 3e-4 | |
| - **Batch Size:** 4 (effective 8 with gradient accumulation) | |
| - **Warmup Steps:** 5 | |
| - **Weight Decay:** 0.01 | |
| - **Precision:** FP16 (mixed precision training) | |
| - **Training Time:** ~1.5 minutes on a 12 GB GPU | |
| # Loss: | |
| - **Training:** 1.0099 (epoch 1) → 0.3479 (epoch 2) | |
| - **Validation:** 1.0176 (epoch 1, best) → 1.0491 (epoch 2) | |
| # Performance | |
| - **Metrics:** Validation loss (best: 1.0176) | |
| - **Qualitative Evaluation:** Generates concise, keyword-focused summaries with good coherence (e.g., "The concert featured a famous singer" for keyword "singer"). | |
| # Intended Use | |
| - Purpose: Summarize short texts (e.g., news snippets, reports) based on a user-specified keyword. | |
| - Use Case: Quick summarization for journalists, researchers, or content creators needing keyword-driven insights. | |
| - Out of Scope: Not designed for long documents (>128 tokens) or abstractive summarization without keywords. | |
| # Usage Instructions | |
| ``` | |
| Requirements | |
| Python 3.8+ | |
| Libraries: transformers, torch, pandas | |
| GPU with FP16 support (e.g., NVIDIA with ~1.5 GB VRAM free) | |
| ``` | |
| # Example Code | |
| ```python | |
| from transformers import T5ForConditionalGeneration, T5Tokenizer | |
| # Load model and tokenizer | |
| model = T5ForConditionalGeneration.from_pretrained("./t5_summary_finetuned_final_fp16").to("cuda") | |
| tokenizer = T5Tokenizer.from_pretrained("./t5_summary_finetuned_final_fp16") | |
| # Generate summary | |
| text = "A new laptop was released with a fast processor and sleek design. It’s popular among gamers." | |
| keyword = "processor" | |
| input_text = f"{text} Keyword: {keyword}" | |
| inputs = tokenizer(input_text, max_length=128, truncation=True, padding="max_length", return_tensors="pt").to("cuda") | |
| outputs = model.generate(input_ids=inputs["input_ids"], attention_mask=inputs["attention_mask"].to(torch.float16), max_length=128, num_beams=4, early_stopping=True, no_repeat_ngram_size=2) | |
| summary = tokenizer.decode(outputs[0], skip_special_tokens=True) | |
| print(summary) # Expected: "The new laptop has a fast processor." | |
| ``` | |