🎓 SmolLM2-360M-Assignment-Metadata-Extractor (GGUF)

This is a highly specialized, lightweight (360M parameter) model fine-tuned specifically to extract student metadata from chaotic, noisy assignment text and output it as strictly formatted JSON.

It was finetuned and converted to 4-bit GGUF format using Unsloth for maximum CPU/GPU efficiency and rapid deployment via Ollama or llama.cpp.

Github Repo: https://github.com/nmdra/Assignment-Metadata-Extractor

📌 Model Capabilities

Unlike generic LLMs, this model has been purposefully overfit on a highly mutated dataset to act as a Zero-Shot Data Extractor. It excels at:

• Noise Filtering: Completely ignoring conversational filler, apologies, word counts, formatting artifacts, and academic instructions.
• Handling Chaos: Robust against typos (e.g., "Stuednt No"), varied capitalization, and unpredictable line breaks.
• Strict JSON Output: Trained to output ONLY a valid JSON object with zero conversational preamble (no "Here is the JSON...").

Expected Output Schema

The model will exclusively output data in the following JSON structure:

{
  "student_number": "...",
  "student_name": "...",
  "assignment_number": "..."
}

---

🚀 Deployment & Usage

Because this model was trained with a specific instruction template, it performs best when wrapped in an environment that enforces Temperature 0 and matches the training prompt.

Method 1: Using Ollama (Recommended for standard usage)

Create a Modelfile with the following configuration to enforce the correct prompt template and prevent creativity:

FROM hf.co/nimendraai/SmolLM2-360M-Assignment-Metadata-Extractor:Q4_K_M

TEMPLATE """### Instruction:
Extract student info as JSON from the following text.

### Input:
{{ .Prompt }}

### Response:
"""

SYSTEM """
You are a precise student assignment data extractor.
Output ONLY a valid JSON object. No explanation. No extra text. No markdown.
Always output exactly: {"student_number":"...","student_name":"...","assignment_number":"..."}
"""

PARAMETER temperature 0
PARAMETER stop "}"

Build and Run:

ollama create json-extractor -f Modelfile
ollama run json-extractor "Course: CS101 \n Stuednt No=20210088 \n Full Nme: Nimal Silva \n HW No.-03 \n Please grade fairly!"

Method 2: Python using Outlines (For bulletproof JSON validation)

For production environments where json.JSONDecodeError is entirely unacceptable, use this model with outlines and llama-cpp-python to structurally constrain the output tokens.

import outlines
from pydantic import BaseModel

class StudentExtraction(BaseModel):
    student_number: str
    student_name: str
    assignment_number: str

# Load the GGUF model
model = outlines.models.llamacpp(
    "hf.co/nimendraai/SmolLM2-360M-Assignment-Metadata-Extractor:Q4_K_M",
    device="cpu" # or "cuda"
)

# Constrain the generator to the Pydantic schema
generator = outlines.generate.json(model, StudentExtraction)

# Format the prompt exactly as trained
prompt = (
    "### Instruction:\nExtract student info as JSON from the following text.\n\n"
    "### Input:\nStu. ID: 20210088 | Full Name: Nimal Silva | HW-3\n\n"
    "### Response:\n"
)

result = generator(prompt)
print(result.model_dump_json())

---

🧠 Training Details

• Base Model: HuggingFaceTB/SmolLM2-360M
• Dataset: 1,250 highly varied synthetic examples containing realistic human errors, markdown noise, and distractor text.
• Epochs: 5 (Optimized to achieve a loss < ~0.40 to prevent hallucinations).
• Framework: Trained efficiently using LoRA (Low-Rank Adaptation) via Unsloth.
• Quantization: Exported to Q4_K_M GGUF format to reduce memory footprint to ~270MB.

---

This was trained 2x faster with Unsloth