MyVillage Activity Generation v1.3.5 — Qwen2.5-0.5B — 427 Checkpoint

Fine-tuned from Qwen/Qwen2.5-0.5B-Instruct for MyVillage next-activity generation.

Training

• Training examples: 427
• Training method: supervised fine-tuning with QLoRA
• Starting checkpoint: clean Qwen/Qwen2.5-0.5B-Instruct
• Epochs: 4
• Learning rate: 0.0001
• LoRA rank: 16
• Target context length: 1024

The repository name uses 325 as the experiment/checkpoint label. The validated training dataset used by this notebook contains 427 examples.

Input

The model expects:

• village goal
• previous activity title
• knowledge submission
• one wisdom object containing book name, book type, chapter title, and content

Output

The model is trained to return JSON with:

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

Intended behavior

The knowledge submission is treated as the strongest evidence of the learner's current state.

The model is trained to avoid repeating work already demonstrated, handle vague submissions conservatively, avoid using the word but as a shortcut, and use the village goal as long-term context rather than blindly forcing the immediate activity toward it.

Important limitation

This is an experimental small-model checkpoint trained on a small synthetic dataset. Behavioral quality should be evaluated on held-out MyVillage scenarios before production use.