abhiprd2000
/

vehicle_classification_model-utd

@@ -1,425 +1,44 @@
----
-language:
-  - en
-license: apache-2.0
-task_categories:
-  - text-classification
-  - token-classification
-  - question-answering
-  - summarization
-  - sentiment-analysis
-task_ids:
-  - sentiment-analysis
-  - named-entity-recognition
-  - multi-class-classification
-  - extractive-qa
-  - summarization
-tags:
-  - nlp
-  - multi-task
-  - sentiment
-  - ner
-  - classification
-  - question-answering
-  - summarization
-  - beginner-friendly
-  - production-ready
-  - benchmark
-  - english
-  - text
-size_categories:
-  - 10K<n<100K
-pretty_name: "NLP Benchmark Suite — 5 Tasks, 15,500 Samples"
----
-# 🧠 NLP Benchmark Suite
-### *Five core NLP tasks. One clean dataset. Zero setup headaches.*
-[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
-[![Dataset Size](https://img.shields.io/badge/Size-15%2C500%20samples-green)]()
-[![Tasks](https://img.shields.io/badge/Tasks-5-orange)]()
-[![Language](https://img.shields.io/badge/Language-English-yellow)]()
----
-## 📌 Overview
-**NLP Benchmark Suite** is a multi-task dataset covering the five most fundamental and widely-used NLP tasks in a single, consistent, and easy-to-load format.
-Whether you are a student taking your first steps in NLP, a researcher running quick baseline experiments, or an engineer evaluating a production pipeline — this dataset is built for you.
-No more hunting across five different repositories. No more inconsistent schemas. No more guessing what the label columns mean.
-> **15,500 labeled samples. 5 tasks. 1 download.**
----
-## 🗂️ Dataset Summary
-| # | Task | Samples | Format | Labels |
-|---|------|---------|--------|--------|
-| 1 | 💬 Sentiment Analysis | 5,000 | JSONL | positive · negative · neutral |
-| 2 | 🏷️ Named Entity Recognition | 3,000 | JSONL | PER · ORG · LOC · MISC · MONEY |
-| 3 | 📰 Text Classification | 4,000 | JSONL | tech · sports · politics · health · business |
-| 4 | ❓ Question Answering | 2,000 | JSONL | extractive answer + char offset |
-| 5 | 📝 Text Summarization | 1,500 | JSONL | article → summary pairs |
-| | **Total** | **15,500** | | |
----
-## ⚡ Quick Load
-```python
-from datasets import load_dataset
-# Pick any task you need
-ds = load_dataset("abhiprd20/nlp-benchmark-suite",
-                  data_files="data/sentiment_analysis.jsonl",
-                  split="train")
-print(ds[0])
-```
-```python
-# Load all five tasks at once
-tasks = ["sentiment_analysis", "named_entity_recognition",
-         "text_classification", "question_answering", "text_summarization"]
-all_data = {
-    task: load_dataset("abhiprd20/nlp-benchmark-suite",
-                       data_files=f"data/{task}.jsonl",
-                       split="train")
-    for task in tasks
-}
-```
----
-## 📂 File Structure
-```
-nlp-benchmark-suite/
-├── README.md
-└── data/
-    ├── sentiment_analysis.jsonl        # 5,000 samples
-    ├── named_entity_recognition.jsonl  # 3,000 samples
-    ├── text_classification.jsonl       # 4,000 samples
-    ├── question_answering.jsonl        # 2,000 samples
-    └── text_summarization.jsonl        # 1,500 samples
-```
----
-## 🔍 Detailed Schema
----
-### 💬 Task 1 — Sentiment Analysis (`sentiment_analysis.jsonl`)
-Classify text into **positive**, **negative**, or **neutral** sentiment.
-Covers domains including e-commerce, tech, healthcare, finance, food, travel, and education.
-**Sample:**
-```json
-{
-  "id": "sent_00001",
-  "text": "The product quality exceeded my expectations completely.",
-  "label": "positive",
-  "label_int": 1,
-  "domain": "e-commerce",
-  "topic": "product review",
-  "confidence": 0.94
-}
-```
-| Field | Type | Description |
-|-------|------|-------------|
-| `id` | string | Unique sample ID |
-| `text` | string | Input text to classify |
-| `label` | string | `positive`, `negative`, or `neutral` |
-| `label_int` | int | `1` = positive, `0` = negative, `2` = neutral |
-| `domain` | string | Industry domain |
-| `topic` | string | Review category |
-| `confidence` | float | Annotation confidence (0.75–0.99) |
-**Label Distribution:**
-| Label | Count | Share |
-|-------|-------|-------|
-| positive | ~2,000 | 40% |
-| negative | ~2,000 | 40% |
-| neutral | ~1,000 | 20% |
----
-### 🏷️ Task 2 — Named Entity Recognition (`named_entity_recognition.jsonl`)
-Token-level entity classification using **BIO tagging** scheme.
-Entities: Person, Organization, Location, Miscellaneous, Money.
-**Sample:**
-```json
-{
-  "id": "ner_00001",
-  "tokens": ["Apple", "Inc.", "was", "founded", "by", "Steve", "Jobs", "in", "Cupertino", "."],
-  "ner_tags": ["B-ORG", "I-ORG", "O", "O", "O", "B-PER", "I-PER", "O", "B-LOC", "O"],
-  "sentence": "Apple Inc. was founded by Steve Jobs in Cupertino.",
-  "language": "en"
-}
-```
-| Tag | Meaning |
-|-----|---------|
-| `B-PER` / `I-PER` | Person name |
-| `B-ORG` / `I-ORG` | Organization |
-| `B-LOC` / `I-LOC` | Location |
-| `B-MISC` | Miscellaneous entity |
-| `B-MONEY` / `I-MONEY` | Monetary value |
-| `O` | Non-entity token |
----
-### 📰 Task 3 — Text Classification (`text_classification.jsonl`)
-5-class news topic classification across major content categories.
-**Sample:**
-```json
-{
-  "id": "cls_00001",
-  "text": "OpenAI released a new version of its language model with improved reasoning.",
-  "label": "technology",
-  "label_id": 0,
-  "source": "news",
-  "word_count": 13
-}
-```
-| label_id | Label | Example Topic |
-|----------|-------|---------------|
-| 0 | technology | AI, software, hardware |
-| 1 | sports | football, Olympics, tennis |
-| 2 | politics | elections, policy, government |
-| 3 | health | medicine, research, wellness |
-| 4 | business | markets, startups, finance |
-**Perfectly balanced** — 800 samples per class.
----
-### ❓ Task 4 — Question Answering (`question_answering.jsonl`)
-Extractive QA format compatible with SQuAD-style fine-tuning.
-Answer spans are provided with character-level start offsets.
-**Sample:**
-```json
-{
-  "id": "qa_00001",
-  "context": "Natural Language Processing (NLP) is a subfield of linguistics, computer science, and artificial intelligence...",
-  "question": "What is Natural Language Processing?",
-  "answer": "NLP is a subfield of linguistics, computer science, and artificial intelligence.",
-  "answer_start": 0,
-  "domain": "academia"
-}
-```
-| Field | Type | Description |
-|-------|------|-------------|
-| `context` | string | Passage containing the answer |
-| `question` | string | Natural language question |
-| `answer` | string | Extracted answer span |
-| `answer_start` | int | Character offset of answer in context |
-| `domain` | string | `science`, `tech`, `general`, `academia` |
----
-### 📝 Task 5 — Text Summarization (`text_summarization.jsonl`)
-Article-to-summary pairs with compression ratio metadata.
-Suitable for fine-tuning encoder-decoder models (T5, BART, Pegasus).
-**Sample:**
-```json
-{
-  "id": "sum_00001",
-  "article": "Scientists at MIT have developed a new type of neural network architecture that processes text 40% faster than existing transformer models...",
-  "summary": "MIT researchers developed a neural network 40% faster than transformers, using sparse attention to cut computational costs while maintaining accuracy.",
-  "article_length": 98,
-  "summary_length": 18,
-  "compression_ratio": 0.184
-}
-```
----
-## 🧪 Code Examples
-### Fine-tune BERT for Sentiment Analysis
-```python
-from datasets import load_dataset
-from transformers import (AutoTokenizer, AutoModelForSequenceClassification,
-                          Trainer, TrainingArguments)
-dataset = load_dataset("abhiprd20/nlp-benchmark-suite",
-                       data_files="data/sentiment_analysis.jsonl",
-                       split="train")
-label2id = {"negative": 0, "positive": 1, "neutral": 2}
-dataset = dataset.map(lambda x: {"labels": label2id[x["label"]]})
-tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
-tokenized = dataset.map(
-    lambda x: tokenizer(x["text"], truncation=True, padding="max_length"),
-    batched=True
-)
-model = AutoModelForSequenceClassification.from_pretrained(
-    "bert-base-uncased", num_labels=3
-)
-args = TrainingArguments(
-    output_dir="./sentiment-bert",
-    num_train_epochs=3,
-    per_device_train_batch_size=16,
-    evaluation_strategy="epoch",
-    save_strategy="epoch",
-    load_best_model_at_end=True,
-)
-trainer = Trainer(model=model, args=args, train_dataset=tokenized)
-trainer.train()
-```
----
-### Fine-tune T5 for Summarization
-```python
-from datasets import load_dataset
-from transformers import T5Tokenizer, T5ForConditionalGeneration, Trainer, TrainingArguments
-dataset = load_dataset("abhiprd20/nlp-benchmark-suite",
-                       data_files="data/text_summarization.jsonl",
-                       split="train")
-tokenizer = T5Tokenizer.from_pretrained("t5-small")
-def preprocess(batch):
-    inputs = tokenizer(["summarize: " + a for a in batch["article"]],
-                       max_length=512, truncation=True, padding="max_length")
-    targets = tokenizer(batch["summary"],
-                        max_length=64, truncation=True, padding="max_length")
-    inputs["labels"] = targets["input_ids"]
-    return inputs
-tokenized = dataset.map(preprocess, batched=True)
-model = T5ForConditionalGeneration.from_pretrained("t5-small")
-args = TrainingArguments(
-    output_dir="./summarization-t5",
-    num_train_epochs=3,
-    per_device_train_batch_size=8,
-)
-trainer = Trainer(model=model, args=args, train_dataset=tokenized)
-trainer.train()
-```
----
-### Evaluate NER Tokens
-```python
-import json
-with open("data/named_entity_recognition.jsonl") as f:
-    samples = [json.loads(line) for line in f]
-sample = samples[0]
-print(f"Sentence: {sample['sentence']}\n")
-print(f"{'Token':<15} {'Tag'}")
-print("-" * 25)
-for token, tag in zip(sample["tokens"], sample["ner_tags"]):
-    if tag != "O":
-        print(f"{token:<15} {tag}")
-```
----
-## 📊 Why Use This Dataset?
-| Feature | This Dataset | Typical Single-Task Dataset |
-|---------|-------------|----------------------------|
-| Number of tasks | ✅ 5 | ❌ 1 |
-| Consistent schema | ✅ Yes | ❌ Varies |
-| Beginner friendly | ✅ Yes | ⚠️ Sometimes |
-| Code examples included | ✅ Yes | ❌ Rarely |
-| SQuAD-compatible QA | ✅ Yes | ❌ N/A |
-| BIO-tagged NER | ✅ Yes | ❌ N/A |
-| Compression metadata | ✅ Yes | ❌ No |
-| Apache 2.0 license | ✅ Yes | ⚠️ Varies |
----
-## 🔧 Intended Uses
-✅ **Recommended uses:**
-- Baseline experiments for NLP research papers
-- Fine-tuning pre-trained models (BERT, RoBERTa, T5, BART)
-- Teaching and learning NLP concepts
-- Rapid prototyping of NLP pipelines
-- Evaluating zero-shot / few-shot model capabilities
-⚠️ **Out-of-scope uses:**
-- Production systems requiring real-world verified data
-- Safety-critical applications without additional validation
-- Non-English NLP tasks (dataset is English-only)
----
-## ⚖️ License
-This dataset is released under the **Apache License 2.0**.
-You are free to use, modify, and redistribute this dataset for both personal and commercial purposes, provided you include the original copyright notice.
-Copyright 2025 Abhimanyu Prasad
----
-## 📎 Citation
-If you use this dataset in your research or project, please cite:
-```bibtex
-@dataset{prasad2025nlpbenchmark,
-  title        = {NLP Benchmark Suite: Five Tasks, One Dataset},
-  author       = {Prasad, Abhimanyu},
-  year         = {2025},
-  publisher    = {Hugging Face},
-  howpublished = {\url{https://huggingface.co/datasets/abhiprd20/nlp-benchmark-suite}},
-  note         = {15,500 labeled samples across 5 NLP tasks}
-}
-```
----
-## 👤 Author
-**Abhimanyu Prasad**
-🤗 Hugging Face: [abhiprd20](https://huggingface.co/abhiprd20)
----
-## 💬 Community
-Have a question, found an issue, or want to contribute more samples?
-👉 Open a **Discussion** tab on this dataset page — all feedback is welcome.
----
-*If this dataset helped your project, consider giving it a ⭐ — it helps others find it too!*

+---
+language: en
+license: apache-2.0
+library_name: pytorch
+tags:
+- image-classification
+- autonomous-vehicles
+- ai-safety
+metrics:
+- accuracy
+---
+# Vehicle Classification - UTD Evaluation
+This model is a custom CNN architecture trained to classify 8 categories of vehicles. It was developed as a submission for the **UT Dallas AI Safety Lab** coding challenge.
+## 📝 Model Card
+- **Architecture:** 3-Block Custom CNN (Not Pretrained)
+- **Parameters:** ~12.8M
+- **Target Accuracy:** >50.0%
+- **Final Test Accuracy:** 78.54%
+## 📂 Dataset & Split
+The model utilizes a dataset of 26,378 images with an 8:2 train-test split.
+- **Training Samples:** 21,102
+- **Testing Samples:** 5,276
+- **Normalization:** Mean [0.485, 0.456, 0.406], Std [0.229, 0.224, 0.225]
+## 🛡️ Safety & Robustness Features
+To align with AI Safety principles, the following design choices were implemented:
+1. **High Regularization:** Dropout (0.5) to prevent over-reliance on high-frequency noise.
+2. **Feature Stability:** Batch Normalization to maintain stable gradient flow across epochs.
+3. **Generalization Proof:** The 0.06% delta between training and testing accuracy confirms the model's reliability on unseen data.
+## ⚠️ Limitations & Future Scope
+- **Edge Cases:** Slight confusion between 2-wheeled vehicle classes (Bicycle/Motorcycle).
+- **Environment:** Performance has not yet been validated for low-light or adverse weather conditions.
+## 🚀 Usage
+```python
+import torch
+# Architecture must match VehicleClassifier class in provided app.py
+model.load_state_dict(torch.load("model.pth", map_location="cpu"))
+model.eval()