---
tags:
- uv-script
- text-classification
- hf-jobs
base_model: LiquidAI/LFM2.5-Encoder-350M
datasets:
- davanstrien/dataset-rows-with-task-categories
pipeline_tag: text-classification
library_name: transformers
---

# dataset-rows-task-classifier

[LiquidAI/LFM2.5-Encoder-350M](https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M) fine-tuned for
multi-label text classification on
[davanstrien/dataset-rows-with-task-categories](https://huggingface.co/datasets/davanstrien/dataset-rows-with-task-categories).

- **Labels (35)**: `audio-classification`, `audio-to-audio`, `automatic-speech-recognition`, `feature-extraction`, `fill-mask`, `image-classification`, `image-feature-extraction`, `image-segmentation`, `image-text-to-text`, `image-to-3d`, `image-to-image`, `image-to-text`, `multiple-choice`, `object-detection`, `question-answering`, `reinforcement-learning`, `robotics`, `sentence-similarity`, `summarization`, `table-question-answering`, `tabular-classification`, `tabular-regression`, `text-classification`, `text-generation`, `text-retrieval`, `text-to-image`, `text-to-speech`, `text-to-video`, `time-series-forecasting`, `token-classification`, … (35 total)
- **Date**: 2026-07-29 13:03 UTC

> [!NOTE]
> This model uses a custom classification head (mean pooling over a backbone without a native sequence-classification class), so loading requires `trust_remote_code=True`. vLLM serving requires a standard architecture.

## Evaluation

| Metric | Value |
|--------|-------|
| f1_micro @ 0.5 | 0.5935 |
| f1_macro @ 0.5 | 0.3879 |
| f1_micro @ tuned | 0.6326 |
| f1_macro @ tuned | 0.5230 |

Per-label decision thresholds tuned on the eval split are stored in
`config.classifier_thresholds`.

**Choosing an operating point**: the stored thresholds maximise per-label F1. For
precision-first use (e.g. auto-applying labels), act only on predictions well above
their threshold — sigmoid probabilities are a usable confidence signal, and filtering
to high-confidence predictions trades coverage for precision. Route the rest to review.

## Usage

```python
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model = AutoModelForSequenceClassification.from_pretrained("davanstrien/dataset-rows-task-classifier", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("davanstrien/dataset-rows-task-classifier", trust_remote_code=True)

inputs = tokenizer("your text here", return_tensors="pt", truncation=True)
probs = torch.sigmoid(model(**inputs).logits)[0]
thresholds = torch.tensor(model.config.classifier_thresholds)  # tuned on validation
labels = [model.config.id2label[i] for i in (probs >= thresholds).nonzero().flatten().tolist()]
print(labels)
```

## Reproduction

Produced on [Hugging Face Jobs](https://huggingface.co/docs/huggingface_hub/guides/jobs) (`gpu`) with the [`train-classifier.py`](https://huggingface.co/datasets/uv-scripts/classification/raw/main/train-classifier.py) recipe from [uv-scripts](https://huggingface.co/uv-scripts). Run it yourself:

```bash
hf jobs uv run --flavor gpu --secrets HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/classification/raw/main/train-classifier.py \
    davanstrien/dataset-rows-with-task-categories davanstrien/dataset-rows-task-classifier --label-column labels
```