Arithmetic Sorl

--- license: apache-2.0 tags: - sorl - arithmetic - interpretability - mechanistic-interpretability - qwen3 --- # Arithmetic SoRL Models Model checkpoints for the **SoRL Arithmetic Interpretability Study**. Small Qwen3 transformers (3L/4H/512d, ~168M params) trained from scratch on integer addition and subtraction, with and without SoRL abstraction tokens. ## Goal Show that SoRL externalizes arithmetic reasoning mechanisms (carry, borrow circuits) as explicit abstraction tokens — observable and intervenable without activation-level tooling. ## Architecture Tiny Qwen3 from random init via `SorlModelWrapper.from_scratch`: ``` hidden_size=512, num_hidden_layers=3, num_attention_heads=4 intermediate_size=2048, vocab_size=151936 ``` ## Experiment subfolders Each subfolder contains a trained model + `train_config.json` + `metrics.json`. | Subfolder | Task | Mode | Abstract Vocab | |---|---|---|---| | `add_baseline` | addition | SFT baseline | 0 | | `add_sorl_abs4` | addition | SoRL v6 | 4 | | `add_sorl_abs8` | addition | SoRL v6 | 8 | | ... | | | | ## Related - Training data: [thoughtworks/arithmetic-sorl-data](https://huggingface.co/datasets/thoughtworks/arithmetic-sorl-data) - Code: [mod_gpt/arithmetic/](https://github.com/fangyuan-ksgk/mod_gpt/tree/amir/arithmetic/arithmetic) - SoRL paper: Yu & Abdullah, "Intention-Level Alignment with Weak Supervision" (2025)