mjpsm/activity-generation-model-v1

Activity Generation Model v1

activity-generation-model-v1 is a fine-tuned version of Qwen2.5-0.5B-Instruct designed to generate the next educational activity for a learner based on three pieces of context:

1. The learner's village goal
2. The learner's previous activity
3. The learner's knowledge submission

The model generates a structured activity containing:

• title
• description
• instructions

The model was developed for the MyVillage learning environment, where learner knowledge submissions can be used to determine an appropriate next activity that builds on demonstrated knowledge while continuing progress toward a larger village goal.

Model Details

• Base Model: Qwen/Qwen2.5-0.5B-Instruct
• Model Type: Causal Language Model
• Approximate Base Model Size: 0.5B parameters
• Fine-Tuning Method: Supervised Fine-Tuning with LoRA
• Frameworks: Transformers, TRL, PEFT
• Primary Language: English
• Task: Educational activity generation
• Output Format: JSON

The LoRA adapter was merged into the base Qwen model after training, producing a standalone model that can be loaded directly with Hugging Face Transformers.

Model Inputs

The model expects three pieces of information.

Village Goal

The larger learning objective that the student's activities should move toward.

Previous Activity

The title of the activity the student completed before submitting their knowledge.

Knowledge Submission

A description of what the student learned, completed, discovered, or demonstrated during the previous activity.

The expected prompt structure is:

Village goal:
{village_goal}

Previous activity:
{previous_activity_title}

Knowledge submission:
{knowledge_submission}

Create the student's next activity.

Model Output

The model is trained to return valid JSON using exactly the following schema:

{
  "title": "activity title",
  "description": "activity description",
  "instructions": "activity instructions"
}

The generated activity should build directly on the learner's knowledge submission while continuing to move the learner toward the village goal.

Training Data

The model was fine-tuned using a synthetic activity-generation dataset containing 1,000 examples.

Each training record follows the general structure:

{
  "input": {
    "village_goal": "...",
    "previous_activity_title": "...",
    "knowledge_submission": "..."
  },
  "output": {
    "title": "...",
    "description": "...",
    "instructions": "..."
  }
}

The training task teaches the model to map learner context to an appropriate next educational activity.

Fine-Tuning Configuration

The model was fine-tuned using LoRA with the following configuration:

Base model: Qwen/Qwen2.5-0.5B-Instruct
Epochs: 3
Maximum sequence length: 1024
Learning rate: 2e-4
Learning-rate scheduler: Cosine
Per-device training batch size: 4
Gradient accumulation steps: 4
Effective batch size: 16
Weight decay: 0.01
LoRA rank: 16
LoRA alpha: 32
LoRA dropout: 0.05
Random seed: 42

LoRA was applied to the following modules:

q_proj
k_proj
v_proj
o_proj
gate_proj
up_proj
down_proj

The best checkpoint was selected using validation loss.

After training, the LoRA adapter was merged into the original Qwen2.5-0.5B-Instruct model using PEFT's merge_and_unload() functionality.

System Prompt

The model was fine-tuned using the following system-level behavior:

You are an educational activity generator for MyVillage.

Your job is to create exactly one logical next learning activity for a student.

You will receive:

1. The goal of the student's village.
2. The title of the student's previous activity.
3. The student's knowledge submission describing what they learned or completed.

Create a new activity that:

- directly builds on the student's knowledge submission;
- moves the student toward the village goal;
- does not simply repeat the previous activity;
- is specific and actionable;
- uses clear student-facing language;
- includes a concrete task or deliverable.

Return valid JSON only.

Return exactly these fields:

{
  "title": "activity title",
  "description": "activity description",
  "instructions": "activity instructions"
}

Do not include markdown.
Do not include commentary.
Do not include additional fields.

How to Use

Install the required packages:

pip install transformers accelerate torch

Then load the model with Hugging Face Transformers.

Replace mjpsm/activity-generation-model-v1 if your Hugging Face repository uses a different name.

import json
import torch

from transformers import AutoModelForCausalLM, AutoTokenizer


MODEL_ID = "mjpsm/activity-generation-model-v1"


tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype="auto",
    device_map="auto",
)

model.eval()

Define the system prompt:

SYSTEM_PROMPT = """You are an educational activity generator for MyVillage.

Your job is to create exactly one logical next learning activity for a student.

You will receive:

1. The goal of the student's village.
2. The title of the student's previous activity.
3. The student's knowledge submission describing what they learned or completed.

Create a new activity that:

- directly builds on the student's knowledge submission;
- moves the student toward the village goal;
- does not simply repeat the previous activity;
- is specific and actionable;
- uses clear student-facing language;
- includes a concrete task or deliverable.

Return valid JSON only.

Return exactly these fields:

{
  "title": "activity title",
  "description": "activity description",
  "instructions": "activity instructions"
}

Do not include markdown.
Do not include commentary.
Do not include additional fields.
"""

Create a generation function:

def generate_activity(
    village_goal,
    previous_activity_title,
    knowledge_submission,
    max_new_tokens=300,
):

    user_message = f"""Village goal:
{village_goal}

Previous activity:
{previous_activity_title}

Knowledge submission:
{knowledge_submission}

Create the student's next activity."""

    messages = [
        {
            "role": "system",
            "content": SYSTEM_PROMPT,
        },
        {
            "role": "user",
            "content": user_message,
        },
    ]

    prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
    )

    inputs = tokenizer(
        prompt,
        return_tensors="pt",
    ).to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            do_sample=False,
            repetition_penalty=1.05,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )

    generated_tokens = outputs[
        0,
        inputs["input_ids"].shape[1]:
    ]

    response = tokenizer.decode(
        generated_tokens,
        skip_special_tokens=True,
    ).strip()

    try:
        return json.loads(response)

    except json.JSONDecodeError:
        return {
            "error": "The model did not return valid JSON.",
            "raw_output": response,
        }

Generate an activity:

activity = generate_activity(
    village_goal="Learn how to build and train machine learning models",

    previous_activity_title="Create a phishing URL dataset",

    knowledge_submission="""
    I created a dataset containing legitimate and phishing URLs.
    I loaded it into Google Colab and cleaned several missing values.
    I have not trained a machine learning model with it yet.
    """,
)

print(
    json.dumps(
        activity,
        indent=2,
        ensure_ascii=False,
    )
)

An output may look similar to:

{
  "title": "Train a Phishing URL Classification Model",
  "description": "Build on your cleaned phishing URL dataset by training a machine learning model that can distinguish between legitimate and phishing URLs.",
  "instructions": "Split your cleaned dataset into training and testing sets. Select a classification algorithm, train the model using the training data, and evaluate its performance on the testing data. Record the model's accuracy and describe what the results tell you."
}

Input and Output Flow

Conceptually, the model performs the following transformation:

Village Goal
      +
Previous Activity
      +
Knowledge Submission
      |
      v
Activity Generation Model
      |
      v
{
  "title": "...",
  "description": "...",
  "instructions": "..."
}

The village goal provides the model with the learner's broader direction, while the previous activity and knowledge submission provide immediate context about the learner's current position.

Intended Use

This model is intended for experimental educational activity generation.

Potential uses include:

• Generating personalized next-step learning activities
• Building learning progression systems
• Supporting activity recommendation workflows
• Generating activities from demonstrated learner knowledge
• Prototyping adaptive learning systems

The model is particularly designed for workflows where a learner completes an activity, submits what they learned, and then receives a new activity that extends that learning.

Limitations

This is an experimental model and should not be treated as an autonomous educational decision-making system.

The model may:

• Generate activities that are too broad or too narrow
• Produce activities that do not perfectly align with a village goal
• Occasionally repeat concepts from a previous activity
• Generate invalid JSON
• Produce instructions that require additional clarification
• Perform less reliably on inputs substantially different from its training distribution

The model was trained using synthetic examples, so performance on real-world learner submissions should be evaluated separately.

Generated activities should be reviewed before being used in high-stakes educational settings.

What This Model Does Not Predict

This version of the model does not predict:

• Activity type
• Estimated activity duration
• Student mastery score
• Activity completion status

Its responsibility is limited to generating:

title
description
instructions

Other attributes can be generated or classified by separate models or application logic.

Future Work

Potential future improvements include:

• Evaluation using real MyVillage knowledge submissions
• Larger and more diverse training datasets
• Automated activity-quality evaluation
• Village-goal alignment scoring
• Knowledge progression evaluation
• Comparison against larger teacher models
• Integration with a separate activity-type classification model
• API deployment for application integration

Base Model

This model was fine-tuned from:

Qwen2.5-0.5B-Instruct

Developed by the Qwen team.

Base model repository:

Qwen/Qwen2.5-0.5B-Instruct

Disclaimer

This model is a research and development prototype. Generated activities should be evaluated for relevance, accuracy, appropriateness, and educational quality before being presented to learners.