README.md

---
library_name: transformers
license: apache-2.0
base_model: distilbert-base-uncased
tags:
- generated_from_trainer
model-index:
- name: help-classifier-v2
  results: []
datasets:
- King-8/help-request-messages-v2
---
# 🤖 Help Classifier Model (v2)

## 🧠 Overview

The **Help Classifier Model (v2)** is a fine-tuned NLP model designed to classify student help requests into meaningful categories within a collaborative learning environment.

This model is part of a larger AI system built for the **Coding in Color (CIC)** ecosystem, supporting students working across domains such as AI development, game development, 2D/3D art, and robotics.

Its primary purpose is to:

* Interpret real student messages
* Identify intent behind help requests
* Route inputs to appropriate downstream systems (e.g., generators, agents)

---

## 🚀 Version Update (v1 → v2)

### 🔹 v1

* Trained on ~100 examples
* Limited generalization
* Struggled with messy or informal input

### 🔹 v2 (Current)

* Trained on **1,000 examples**
* Balanced dataset across all categories
* Strong performance on:

  * informal/slang input
  * mixed tone messages
  * ambiguous phrasing
  * real CIC-style check-ins

👉 v2 significantly improves **accuracy, stability, and real-world usability**

---

## 🧩 Task Definition

**Task Type:** Text Classification

**Input:** Student message
**Output:** One of 5 help categories

---

## 🏷️ Labels

| Label              | Description                                         |
| ------------------ | --------------------------------------------------- |
| `learning_help`    | User is trying to understand a concept or skill     |
| `project_help`     | User needs direction or next steps in a project     |
| `technical_issue`  | Something is broken or not working                  |
| `attendance_issue` | User missed a meeting or needs to catch up          |
| `general_guidance` | User expresses uncertainty, stress, or needs advice |

---

## 🏗️ Model Architecture

* Base Model: distilbert-base-uncased
* Fine-tuned for sequence classification
* Number of labels: 5

---

## ⚙️ Training Configuration

* Epochs: 4
* Learning Rate: 2e-5
* Batch Size: 8
* Weight Decay: 0.01
* Train/Validation Split: 80/10/10

---

## 📊 Training Results

| Epoch | Training Loss | Validation Loss |
| ----- | ------------- | --------------- |
| 1     | 0.552         | 0.512           |
| 2     | 0.111         | 0.122           |
| 3     | 0.032         | 0.077           |
| 4     | 0.025         | 0.064           |

---

## 📈 Performance Summary

* **Low validation loss (~0.06)**
* Strong generalization across unseen inputs
* Stable convergence during training
* Handles:

  * messy/slang text
  * indirect requests
  * multi-layered inputs

---

## 🧪 Example Predictions

**Input:**

```
i missed the meeting and now idk what we’re doing
```

**Output:**

```
attendance_issue
```

---

**Input:**

```
my model works but the predictions are weird and I don’t know why
```

**Output:**

```
technical_issue
```

---

**Input:**

```
I feel like I’m behind and don’t know what to focus on
```

**Output:**

```
general_guidance
```

---

## 🔗 System Integration

This model is integrated into an MCP (Model Context Protocol) system where it acts as:

> **Entry-point classifier for routing student inputs**

Pipeline example:

```
User Input → Help Classifier → (Future: Generator / Summarizer)
```

---

## 🎯 Use Cases

* Help request classification
* Slack/Discord message routing
* Educational AI assistants
* CIC ecosystem tools
* AI agent pipelines

---

## ⚠️ Limitations

* Single-label classification (some messages may contain multiple intents)
* Edge cases may still overlap between categories
* Domain-specific (focused on student tech environments)

---

## 🔮 Future Improvements

* Multi-label classification
* Larger dataset (2,000+ examples)
* Confidence scoring
* Integration with response generation models
* Continuous retraining with real user data

---

## 👤 Author

Created by Kingston Lewis as part of the Coding in Color program for the AI Dev team.

---

# help-classifier-v2

This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on the King-8/help-request-messages-v2 dataset.
It achieves the following results on the evaluation set:
- Loss: 0.0643


### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- num_epochs: 4
- mixed_precision_training: Native AMP


### Training results

| Training Loss | Epoch | Step | Validation Loss |
|:-------------:|:-----:|:----:|:---------------:|
| 0.5524        | 1.0   | 88   | 0.5124          |
| 0.1114        | 2.0   | 176  | 0.1221          |
| 0.0324        | 3.0   | 264  | 0.0771          |
| 0.0249        | 4.0   | 352  | 0.0643          |


### Framework versions

- Transformers 5.0.0
- Pytorch 2.10.0+cpu
- Datasets 4.0.0
- Tokenizers 0.22.2