thoughtworks/arithmetic-sorl · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Modellquelle
- Art
- Other
- Dateigröße
- 150–621 MB
- Aktualisiert
- 22. Sept. 2026
- Verleger
- @thoughtworks
- Quelle
- Hugging Face
Quellenauszug
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
Quellen
1 Quelle
Quellen
1 QuelleModellartefakte
64 Artefakte
Modellartefakte
64 Artefakteadd_sub_baseline_100K_1L2H256d/model.safetensors
safetensors · 300 MB · SHA-256 647cf0cb4128…32c8
add_sub_baseline_100K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 df040a739091…41b7
add_sub_baseline_100K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 dc9a0ac6c9a2…9d0d
add_sub_baseline_10K_1L2H256d/model.safetensors
safetensors · 300 MB · SHA-256 aa05a1cd94e5…4c12
add_sub_baseline_10K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 b5bbfcb52eff…3c18
add_sub_baseline_10K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 8a71d76648b9…73e7
add_sub_baseline_10K/model.safetensors
safetensors · 620 MB · SHA-256 93cfd7c10000…5e0b
add_sub_baseline_25K_1L2H256d/model.safetensors
safetensors · 300 MB · SHA-256 3fb5c0fa9419…c3c1
add_sub_baseline_25K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 bc1136aa94b4…85dd
add_sub_baseline_25K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 2da20f6f58ed…45ab
add_sub_baseline_25K/model.safetensors
safetensors · 620 MB · SHA-256 9d7ab48f0344…daaa
add_sub_baseline_50K_1L2H256d/model.safetensors
safetensors · 300 MB · SHA-256 e7a47a187ec4…dd06
add_sub_baseline_50K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 52d5fc071908…9219
add_sub_baseline_50K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 d6aa50140e3d…1dd5
add_sub_baseline_50K/model.safetensors
safetensors · 620 MB · SHA-256 e58411581008…cafa
add_sub_sorl_v1_abs1_K1_100K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 276c4ccb51fd…1d38
add_sub_sorl_v1_abs1_K1_100K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 6655d80df0ed…057f
add_sub_sorl_v1_abs1_K1_10K/model.safetensors
safetensors · 620 MB · SHA-256 b657f0d2d656…3f91
add_sub_sorl_v1_abs10_100K/model.safetensors
safetensors · 620 MB · SHA-256 7efc43879466…7d91
add_sub_sorl_v1_abs10_10K/model.safetensors
safetensors · 620 MB · SHA-256 324ce6c56d72…5205
add_sub_sorl_v1_abs10_25K/model.safetensors
safetensors · 620 MB · SHA-256 a787d3e4f23b…1578
add_sub_sorl_v1_abs10_50K/model.safetensors
safetensors · 620 MB · SHA-256 45c37fed66be…b81a
add_sub_sorl_v1_abs10_K1_100K_1L2H256d/model.safetensors
safetensors · 300 MB · SHA-256 13c22ddce9bf…7a0f
add_sub_sorl_v1_abs10_K1_100K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 eaca29567d0d…7937
add_sub_sorl_v1_abs10_K1_100K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 c61e97d3aa07…4154
add_sub_sorl_v1_abs10_K1_100K/model.safetensors
safetensors · 620 MB · SHA-256 b4c0077633c8…56a6
add_sub_sorl_v1_abs10_K1_10K/model.safetensors
safetensors · 620 MB · SHA-256 47ed228ea959…7aed
add_sub_sorl_v1_abs10_K1_25K_1L2H256d/model.safetensors
safetensors · 300 MB · SHA-256 47628d01fe60…d2b0
add_sub_sorl_v1_abs10_K1_25K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 abedcbc3a527…d329
add_sub_sorl_v1_abs10_K1_25K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 dcb5ae69edf2…f6a6
add_sub_sorl_v1_abs10_K1_25K/model.safetensors
safetensors · 620 MB · SHA-256 17d304a942f9…7635
add_sub_sorl_v1_abs10_K1_50K/model.safetensors
safetensors · 620 MB · SHA-256 8ceaee5d7548…d3d4
add_sub_sorl_v1_abs100_100K/model.safetensors
safetensors · 621 MB · SHA-256 0e4b24a426ca…1faf
add_sub_sorl_v1_abs100_K1_100K/model.safetensors
safetensors · 621 MB · SHA-256 e2a8948743a4…b855
add_sub_sorl_v1_abs16_100K/model.safetensors
safetensors · 620 MB · SHA-256 8af91a334202…a8a5
add_sub_sorl_v1_abs16_K1_100K/model.safetensors
safetensors · 620 MB · SHA-256 2653ee350cd3…b8af
add_sub_sorl_v1_abs2_100K/model.safetensors
safetensors · 620 MB · SHA-256 a3139180f867…a489
add_sub_sorl_v1_abs2_K1_100K_1L2H256d/model.safetensors
safetensors · 300 MB · SHA-256 5cae09e943cf…8617
add_sub_sorl_v1_abs2_K1_100K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 04a6c769c661…b3d4
add_sub_sorl_v1_abs2_K1_100K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 63796e8e977b…ef98
add_sub_sorl_v1_abs2_K1_100K/model.safetensors
safetensors · 620 MB · SHA-256 3a3de8400fb4…6384
add_sub_sorl_v1_abs2_K1_10K/model.safetensors
safetensors · 620 MB · SHA-256 556319ae5991…d6ee
add_sub_sorl_v1_abs20_100K/model.safetensors
safetensors · 620 MB · SHA-256 bb9ee27e51d6…a5db
add_sub_sorl_v1_abs20_K1_100K_1L2H256d/model.safetensors
safetensors · 301 MB · SHA-256 90730763dca7…ec88
add_sub_sorl_v1_abs20_K1_100K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 950f9d8cc080…3cc7
add_sub_sorl_v1_abs20_K1_100K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 f343bc391b31…8204
add_sub_sorl_v1_abs20_K1_100K/model.safetensors
safetensors · 620 MB · SHA-256 00ab15b3dfda…46da
add_sub_sorl_v1_abs20_K1_10K/model.safetensors
safetensors · 620 MB · SHA-256 22e372b9ada9…11d6
add_sub_sorl_v1_abs30_100K_1L2H256d/model.safetensors
safetensors · 301 MB · SHA-256 09d98deafce7…3295
add_sub_sorl_v1_abs30_100K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 31511e238e55…ae98
add_sub_sorl_v1_abs30_100K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 dcb423af7c9b…04b3
add_sub_sorl_v1_abs30_100K/model.safetensors
safetensors · 620 MB · SHA-256 e953953ca0d0…ebed
add_sub_sorl_v1_abs30_10K/model.safetensors
safetensors · 620 MB · SHA-256 688cc5ac3071…4636
add_sub_sorl_v1_abs30_50K/model.safetensors
safetensors · 620 MB · SHA-256 7d4331d220d9…ba47
add_sub_sorl_v1_abs30_K1_100K_1L2H256d/model.safetensors
safetensors · 301 MB · SHA-256 3d7724247590…20db
add_sub_sorl_v1_abs30_K1_100K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 fc66795e1f95…0d3b
add_sub_sorl_v1_abs30_K1_100K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 3a489deabd1d…0059
add_sub_sorl_v1_abs30_K1_100K/model.safetensors
safetensors · 620 MB · SHA-256 3e8b4c1e17d7…0ef8
add_sub_sorl_v1_abs30_K1_10K_1L2H256d/model.safetensors
safetensors · 301 MB · SHA-256 78a51dafce3a…5476
add_sub_sorl_v1_abs30_K1_10K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 dc844b7e06f9…834b
add_sub_sorl_v1_abs30_K1_10K_2L1H128d/model.safetensors
safetensors · 150 MB · SHA-256 25d4c381a7ed…9334
add_sub_sorl_v1_abs30_K1_10K/model.safetensors
safetensors · 620 MB · SHA-256 a4a7a8e2c5ce…8749
add_sub_sorl_v1_abs30_K1_25K_1L2H256d/model.safetensors
safetensors · 301 MB · SHA-256 c79d62fba649…b025
add_sub_sorl_v1_abs30_K1_25K_1L3H510d/model.safetensors
safetensors · 605 MB · SHA-256 76f3dde8b6ba…1d1e
Quellenauszüge
3 Auszüge
Quellenauszüge
3 Auszüge--- 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)
Source context: 0 downloads · 2 likes · Repo thoughtworks/arithmetic-sorl
Source context: 0 downloads · 2 likes · Repo thoughtworks/arithmetic-sorl
