activity-generation-v1.3

activity-generation-v1.3 is a fine-tuned version of Qwen/Qwen2.5-0.5B-Instruct built for the MyVillage ecosystem.

Its role is to generate the smallest useful next activity from four pieces of context:

• village_goal
• previous_activity_title
• knowledge_submission
• one wisdom object

The model is designed for micro-progression, not full curriculum generation.

One activity. One primary action. One small progression.

Model Details

• Model: activity-generation-v1.3
• Base model: Qwen/Qwen2.5-0.5B-Instruct
• Training: Supervised Fine-Tuning (SFT)
• Fine-tuning method: LoRA
• Adapter merge: Yes; the trained LoRA adapter was merged back into the base model
• Training data: approximately 1.2K synthetic SFT examples
• Framework: Hugging Face Transformers + PEFT
• Primary task: MyVillage next-activity generation
• Language: English

What Is MyVillage?

MyVillage is a community-centered learning ecosystem built around the idea that communities already contain knowledge worth preserving, sharing, and learning from.

Villagers work toward goals through activities, submit what they learned or completed, contribute wisdom, and use that context to continue progressing.

For this model:

Village Goal
    +
Previous Activity
    +
Knowledge Submission
    +
Wisdom
    ↓
Smallest useful next activity

The model is not intended to generate full curricula, lesson plans, or large assignments.

Intended Use

The model generates one short next activity after a villager completes or attempts a previous activity.

It is designed to:

• generate one small next step
• keep tasks concise
• avoid multi-step assignments
• use wisdom implicitly
• respond appropriately to vague or incomplete knowledge submissions
• ask for clarification or evidence when there is not enough information to progress safely

Input Format

{
  "village_goal": "...",
  "previous_activity_title": "...",
  "knowledge_submission": "...",
  "wisdom": {
    "chapterTitle": "...",
    "bookName": "...",
    "bookTypeName": "...",
    "content": "..."
  }
}

Output Format

The model is trained to predict exactly:

{
  "activity_title": "...",
  "activity_description": "...",
  "activity_instructions": "..."
}

The model is not trained to predict dataset metadata such as:

• department
• domain
• activity type
• difficulty
• submission quality
• messy type
• row IDs
• source metadata

Those fields were used only for dataset generation, balancing, debugging, and analysis.

Output Behavior

Activity title

Should be short, direct, and describe the next micro-progression.

Activity description

Should briefly explain the activity itself.

The model was trained to avoid phrasing such as:

The learner will...
The student will...
The villager will...
The user will...

Activity instructions

Should:

• contain one primary action
• normally be one sentence
• never become a large multi-step assignment
• avoid unnecessary tool usage
• request clarification when the knowledge submission is insufficient

Handling Vague Knowledge Submissions

A major design goal of v1.3 is preventing the model from inventing progress from insufficient evidence.

Example Input

{
  "village_goal": "Understanding how to train and fine-tune AI models",
  "previous_activity_title": "Test Your Classifier Model",
  "knowledge_submission": "Completed",
  "wisdom": {
    "chapterTitle": "Testing With Purpose",
    "bookName": "The Proverbs of Model Building",
    "bookTypeName": "Proverbs",
    "content": "Testing is not a final step; it is a conversation with what you built."
  }
}

Base Model Output

{
  "activity_title": "Testing Your Classifier Model",
  "activity_description": "I tested my classifier model by training it on a dataset and evaluating its performance. I also fine-tuned the model using cross-validation techniques.",
  "activity_instructions": "I will now test my classifier model again to ensure its accuracy and reliability. Please provide feedback on the results."
}

The base model invented details that were not provided, including training, cross-validation, and evaluation behavior.

Fine-Tuned v1.3 Output

{
  "activity_title": "Show Your Classifier Model",
  "activity_description": "Describe your classifier model test result.",
  "activity_instructions": "What did you get?"
}

The fine-tuned model recognized that Completed did not provide enough evidence for a technical next step and instead asked for clarification.

Training Data

The model was trained on approximately 1.2K synthetic examples created specifically for the MyVillage activity-generation task.

The synthetic pipeline used real MyVillage material as reference guidance for style and structure, including:

• village goals
• previous activity titles
• knowledge submissions
• MyVillage wisdom

The dataset intentionally included both clean and messy knowledge submissions so the model could learn clarification behavior.

The training data also covered:

• Coding in Color domains
• broader MyVillage domains
• CREATE, REFLECTION, and RESEARCH activity patterns
• Chronicle, Testimony, Parables, and Proverbs wisdom styles

Training Method

The model was fine-tuned using LoRA on top of Qwen/Qwen2.5-0.5B-Instruct.

LoRA targeted Qwen attention and MLP projection layers including:

q_proj
k_proj
v_proj
o_proj
gate_proj
up_proj
down_proj

After training, the adapter was merged into the base model:

Qwen/Qwen2.5-0.5B-Instruct
        +
LoRA adapter
        ↓
merged activity-generation-v1.3

The merged model can be loaded directly without attaching the adapter separately.

Training Objective

The training pipeline used assistant-only loss masking.

System Prompt       → ignored by loss
User Context        → ignored by loss
Assistant Activity  → trained

This focuses the training signal on the three activity output fields rather than teaching the model to reproduce its inputs.

Evaluation

The merged model was compared against the untouched base model on held-out examples.

Observed improvements included:

• much stronger JSON output consistency
• shorter and more focused activities
• less copying of wisdom
• less repetition of the previous activity
• better handling of vague submissions
• fewer invented details
• stronger micro-progression behavior

The strongest improvement was clarification behavior for vague submissions such as:

Completed
just did the activity and it's done now
started testing game mechanics

The fine-tuned model generally responded by asking the villager to show or describe what happened instead of inventing unseen progress.

Known Limitations

Micro-progression strictness

The model may still occasionally produce a task that is too large.

For example:

Research and experiment with 5 different compression techniques.

A better activity would be:

Research one compression technique that could reduce the accuracy loss.

Best-next-step precision

The model can generate a relevant progression without always selecting the most precise next step.

Unsupported specificity

The model may occasionally introduce details that were not explicitly established by the input.

Small model capacity

The base model is approximately 0.5B parameters, so expectations should match a small specialized model rather than a large general-purpose LLM.

Age agnostic

The model currently receives no explicit age information.

Out-of-Scope Uses

This model should not be relied on for:

• high-stakes educational decisions
• grading
• determining mastery
• psychological assessment
• medical, legal, or financial advice
• full curriculum generation
• fully autonomous progression without application-level validation

Future Work

The next planned stage is Direct Preference Optimization (DPO).

SFT has taught the model the basic task. DPO will focus on preference-level improvements such as:

one small action
vs
multi-step assignment

precise next step
vs
generic next step

supported context
vs
invented details

clarification when needed
vs
fabricated progression

Planned pipeline:

Qwen/Qwen2.5-0.5B-Instruct
        ↓
activity-generation-v1.3 SFT
        ↓
chosen / rejected preference pairs
        ↓
DPO
        ↓
improved activity-generation model

Loading the Merged Model

from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "YOUR_USERNAME/activity-generation-v1.3"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

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

Basic Inference Example

import json
import torch

SYSTEM_PROMPT = """You are MyVillage's activity-generation model.

Generate exactly one small next micro-progression activity.

Rules:
- One activity, one primary action, one small progression.
- Use the village goal, previous activity, knowledge submission, and wisdom.
- If the knowledge submission is vague or insufficient, ask for clarification before progressing.
- Do not quote or explicitly mention the wisdom.
- Return only valid JSON with exactly these keys:
  activity_title
  activity_description
  activity_instructions
"""

payload = {
    "village_goal": "Understanding how to train and fine-tune AI models",
    "previous_activity_title": "Test Your Classifier Model",
    "knowledge_submission": "Completed",
    "wisdom": {
        "chapterTitle": "Testing With Purpose",
        "bookName": "The Proverbs of Model Building",
        "bookTypeName": "Proverbs",
        "content": "Testing is not a final step; it is a conversation with what you built."
    }
}

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": json.dumps(payload, indent=2)}
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
)

device = next(model.parameters()).device
inputs = {k: v.to(device) for k, v in inputs.items()}

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=180,
        do_sample=False,
        repetition_penalty=1.05,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

generated = output[0, inputs["input_ids"].shape[-1]:]

print(
    tokenizer.decode(
        generated,
        skip_special_tokens=True,
    )
)

Attribution

If using this model publicly, please credit:

activity-generation-v1.3
Built for the MyVillage ecosystem
Base model: Qwen/Qwen2.5-0.5B-Instruct

Disclaimer

This is an experimental specialized activity-generation model.

Outputs should be evaluated within the MyVillage product context before production use. The model may still produce irrelevant, overly broad, unsupported, or malformed outputs, especially for inputs outside its training distribution.