---
license: apache-2.0
language:
- en
base_model:
- HuggingFaceTB/SmolLM2-360M
datasets:
- HuggingFaceTB/everyday-conversations-llama3.1-2k
pipeline_tag: text-classification
tags:
- SmolLM2
- conversational
---

# SmolLM2 - Sentence Completeness Detection

This model is a fine-tuning of [SmolLM2](https://huggingface.co/HuggingFaceTB/SmolLM2-360M) on a custom **sentence completeness classification** task.<br />
The goal of the model is to determine whether a given sentence is **complete** or has been **cut off mid-thought** — for example, when a user accidentally presses enter while still typing.

---

## Use Case

This model is ideal for applications where users enter free-form text and you want to:

- Detect accidental sentence cutoffs  
- Improve UX in chatbots, note-taking apps, writing assistants, etc.  
- Prompt users to finish their thought when necessary

---

## Task Definition

Given a single sentence, the model predicts:

- `label=1` → The sentence is **complete**  
- `label=0` → The sentence is **incomplete / cut off**

---

## Example Inputs

| Sentence | Prediction |
|----------|------------|
| `"I think we should meet tomorrow around 3pm at"` | `Incomplete (label=0)` |
| `"Let's catch up after the meeting tomorrow."` | `Complete (label=1)` |
| `"The best thing about this project is"` | `Incomplete (label=0)` |

---

## How to Use

```python
# pip install transformers torch lightning peft torchmetrics
from huggingface_hub import hf_hub_download
from torch import no_grad, sigmoid
import importlib.util
import sys

repo_id = "ZivK/smollm2-end-of-sentence"
model_name = "token_model.ckpt"
model_src_name = "model.py"
checkpoint_path = hf_hub_download(repo_id=repo_id, filename=model_name)
model_src_path = hf_hub_download(repo_id=repo_id, filename=model_src_name)

# Load the model source code, you can also just download model.py
spec = importlib.util.spec_from_file_location("SmolLM", model_src_path)
smollm_model = importlib.util.module_from_spec(spec)
sys.modules["smollm_model"] = smollm_model
spec.loader.exec_module(smollm_model)

device = "cuda" # for GPU usage or "cpu" for CPU usage
label_map = {0: "Incomplete", 1: "Complete"}
model = smollm_model.SmolLM.load_from_checkpoint(checkpoint_path).to(device)
inputs = model.tokenizer("Gravity is", return_tensors="pt").to(device)
model.eval()
with no_grad():
    logits = model(inputs)
    probs = sigmoid(logits)
    prediction = (probs > 0.5).int().item()
    label = label_map[prediction]
    conf = probs.item() if probs.item() > 0.5 else 1 - probs.item()
print(f"Sentence is {label}, Confidence: {conf*100}%")

```

> **Tip:** Use the result to prompt users with _"Do you want to finish your sentence?"_ or similar UX.

---

## Model Details

- **Base model**: [SmolLM2 (360M)](https://huggingface.co/HuggingFaceTB/SmolLM2-360M)  
- **Fine-tuned on**: A custom binary classification dataset of complete/incomplete sentences
- **Labels**:  
  - `LABEL_0`: Incomplete  
  - `LABEL_1`: Complete

---

## Limitations

- Will assume you will not write an incomplete sentence with a dot or question mark in the end, as it will probably classify it as complete even if it's not complete.
- Same limitation as per [SmolLM2's model card](https://huggingface.co/HuggingFaceTB/SmolLM2-360M):<br/><blockquote>SmolLM2 models primarily understand and generate content in English. They can produce text on a variety of topics, but the generated content may not always be factually accurate, logically consistent, or free from biases present in the training data. These models should be used as assistive tools rather than definitive sources of information. Users should always verify important information and critically evaluate any generated content.</blockquote>