---
license: mit
library_name: pytorch
tags:
  - time-series
  - accelerometer
  - ecg
  - sleep
  - nilm
  - capture24
  - ltaf
  - uk-dale
  - sleep-psg
  - classifier
---

# ARTS-RLM Tool Classifiers

PyTorch checkpoints for the time-series classifiers exposed as **tools** to the
recursive agent in *Recursive Agents are Effective Time Series Reasoners*
(ARTS-RLM). Each subset of the benchmark (capture24 / ltaf / sleep_psg /
uk_dale) gets a small specialist classifier the agent can call as a tool while
reasoning over long-context signals.

## Layout

The repo mirrors the on-disk paths the ARTS-RLM configs already point at, so
files can be dropped straight into the project tree.

```
.
├── results/classifier/
│   ├── chronos2/best_classifier.pt          # Capture24 — Chronos-2 encoder + head (active in configs/rlm)
│   ├── oxwearables/best_classifier.pt       # Capture24 — OxWearables baseline (+ results.json)
│   └── dual/
│       ├── best_classifier.pt               # Capture24 — dual-encoder (pretrain)
│       └── finetune/best_classifier.pt      # Capture24 — dual-encoder (finetune)
├── results/sleep_classifier/
│   ├── sleep_stages/best_classifier.pt      # Sleep-PSG — 5-stage sleep classifier
│   └── arousals/best_classifier.pt          # Sleep-PSG — arousal detector (+ results.json)
├── checkpoints/ltaf/
│   ├── rhythm_resnet1d/best_classifier.pt   # LTAF — rhythm (AF/AFL/NSR/…) ResNet1D
│   └── beats_htf/best_classifier.pt         # LTAF — beat (N/V/S) HTF head
└── results/uk_dale_classifier{,_v2,_v2_ft}/best_classifier.pt
                                             # UK-DALE — appliance window classifier
                                             # v1: active eval/RLM configs
                                             # v2 / v2_ft: arts_rlm_qwen35_9b.toml smoke
```

## Which checkpoint do the configs use?

| Subset      | Tool                            | Checkpoint                                                   |
|-------------|---------------------------------|--------------------------------------------------------------|
| capture24   | `har_classifier`                | `results/classifier/chronos2/best_classifier.pt`             |
| ltaf        | `ecg_rhythm`                    | `checkpoints/ltaf/rhythm_resnet1d/best_classifier.pt`        |
| ltaf        | `ecg_beats`                     | `checkpoints/ltaf/beats_htf/best_classifier.pt`              |
| sleep_psg   | `sleep_classifier`              | `results/sleep_classifier/sleep_stages/best_classifier.pt`   |
| sleep_psg   | `sleep_classifier` (arousals)   | `results/sleep_classifier/arousals/best_classifier.pt`       |
| uk_dale     | `uk_dale_appliance_window`      | `results/uk_dale_classifier/best_classifier.pt`              |

The alternative `uk_dale_classifier_v2{,_ft}` and `results/classifier/{oxwearables,dual}` are shipped for the
ablation table.

## Download

Use the ARTS-RLM helper (drops files into the locations the configs expect):

```bash
python scripts/data/download_from_hf.py --dataset classifiers
```

Or pull individual files directly:

```bash
huggingface-cli download nz00shuuuu/arts-rlm-classifiers \
    results/classifier/chronos2/best_classifier.pt \
    --local-dir .
```

## Loading

Each checkpoint is a plain `torch.save({...})` dict. They are loaded inside
`src/models/tools/` via the respective tool's `_load_classifier(...)` helper —
look at `src/models/tools/registry.py` for the wiring.

```python
import torch
state = torch.load("results/classifier/chronos2/best_classifier.pt", map_location="cpu")
```