Fyve-AI

Fyve-AI is a fine-tuned version of Qwen3-4B trained for one specific task: reading a student's broken Python code and responding with a Socratic 3-sentence hint — not the answer.

It is the AI model powering PyFyve, a fully offline Python tutoring application.

---

What It Does

Given a task description, the student's buggy code, and the Python error it produced, the model outputs a JSON object with two fields:

• reasoning — internal diagnosis of what went wrong and why
• hint — exactly 3 sentences following a fixed structure:
  1. Diagnosis — names the specific variable, expression, or construct that caused the error
  2. Rule — states the Python rule that was violated
  3. Directive — starts with "Think about..." or "Consider..." and guides without giving the fix

The model never gives corrected code. It never gives more than 3 sentences. It does not know how to do anything outside this task.

---

Input Format

Task:
<what the student was asked to do>
Code:
<the student's broken code>
Error:
<the Python error message>

---

Output Format

{
  "reasoning": "...",
  "hint": "Sentence 1.\nSentence 2.\nSentence 3."
}

---

Example

Input:

Task:
Create a variable score = 95 and print its value.
Code:
score = 95
print(Score)
Error:
NameError: name 'Score' is not defined at line 2

Output:

{
  "reasoning": "score is defined lowercase but Score (capital S) is used in print. I name both spellings, explain case sensitivity, and direct toward comparing the two usages.",
  "hint": "You defined a variable called score on line 1 but referenced Score on line 2.\nIn Python, variable names are case-sensitive, so score and Score are treated as two completely different identifiers.\nConsider whether the capitalisation of the variable name is consistent between where it was defined and where it is used."
}

---

Training Details

Detail	Value
Base model	Qwen3-4B
Method	QLoRA via Unsloth
Hardware	Google Colab T4 (free tier)
Dataset	555 curated (task, code, error, hint) pairs
Dataset source	Synthetic — generated using Qwen3-30B-A3B as teacher model
Error types covered	SyntaxError, NameError, TypeError, IndexError, KeyError, ValueError, AttributeError, UnboundLocalError, RecursionError, ZeroDivisionError, and more

The training data was generated by a 30B teacher model, manually reviewed for quality, and filtered through a validation pipeline that checks hint structure, sentence count, and semantic rules (e.g. AttributeError on strings must guide toward + or +=, not list conversion).

---

Intended Use

This model is designed exclusively for use inside the PyFyve app. It is not a general-purpose assistant and will produce poor results for tasks outside its training distribution.

It is not designed to:

• Answer general Python questions
• Explain concepts freely
• Write or complete code
• Serve as a chatbot

---

Limitations

• Trained on 555 examples — covers common beginner and intermediate Python errors well, but unusual or advanced errors may produce weaker hints
• No coverage of logic errors (code that runs but produces wrong output)
• Some uncommon syntax patterns (e.g. trailing comma creating a tuple) are outside the training distribution
• The 3-sentence format is enforced by the prompt at inference time — removing the few-shot examples from the prompt degrades output quality significantly

---

Usage with Ollama

This model is distributed as a GGUF file for use with Ollama. The Modelfile in this repository contains the Ollama model definition.

ollama create fyve-ai -f Modelfile

Or use the PyFyve app, which handles setup automatically.

---

License

The fine-tuned weights are released under the same license as the base model: the Apache License.

Please read it before redistributing — it permits research and personal use but has restrictions on commercial use above certain usage thresholds.

---

Citation

If you use this model in research or build on it, please link back to the PyFyve repository.