| --- |
| license: cc-by-4.0 |
| tags: |
| - sparse-autoencoder |
| - interpretability |
| - sonar |
| - automatic-interpretability |
| --- |
| |
| # SONAR SAEs — autointerp results |
|
|
| Automatic interpretability outputs (latent explanations and held-out |
| classifier scores) produced from the scaled-up BatchTopK SAE in: |
|
|
| > **Interpretability of Text Auto-Encoders using Sparse Auto-Encoders: A Sandbox for Interpreting Neuralese.** |
| > Nicky Pochinkov & Jason Rich Darmawan, EACL 2026 (submitted). |
|
|
| ## Protocol |
|
|
| For each SAE latent we: |
|
|
| 1. Selected the top-10 sentences with the highest activation values. |
| 2. Prompted **GPT-OSS-120B** (`openai/gpt-oss-120b`) to generate a |
| short natural-language description of what the latent appears to |
| detect. |
| 3. Re-prompted the same model on a shuffled set of 12 sentences (8 |
| random + 4 top-activating) to classify each as matching or not |
| matching the generated explanation. |
| 4. Reported accuracy / precision / recall / F1 against the held-out |
| labels. |
|
|
| **Caveat (also stated in the paper).** Because explanation generation |
| and evaluation use the same model, high F1 should be read as |
| self-consistency rather than external semantic validation. |
|
|
| ## Files |
|
|
| Each subdirectory is one SAE run id and contains per-latent JSON |
| records of the form: |
|
|
| ```json |
| { |
| "latent_id": 41816, |
| "top_activations": [...], |
| "explanation": "sentences that include a lexical item referring to a cat", |
| "scores": {"f1": 1.0, "precision": 1.0, "recall": 1.0, "accuracy": 1.0}, |
| "heldout_examples": [...] |
| } |
| ``` |
|
|
| ## Related |
|
|
| - [`nickypro/sonar-saes-large`](https://huggingface.co/nickypro/sonar-saes-large) — the SAE these results were generated from |
|
|
