Upload 4 files

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Shiv Shanmugam
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -1,3 +1,88 @@
 ---
 license: mit
+library_name: pytorch
+pipeline_tag: other
+tags:
+  - cad
+  - freecad
+  - agent
+  - imitation-learning
+  - length-generalization
 ---
+
+# Taiga-S1
+
+![Taiga-S1: parts the model built in FreeCAD](assets/cover.png)
+
+**A 1.2M-parameter model that builds CAD parts in FreeCAD.**
+
+*Taiga-S1 is an experiment in whether small, fast decision models can be useful for computer-use agents: a planner decides what to do, and a tiny model handles the step-by-step execution. FreeCAD is the testbed.*
+
+Taiga-S1 is the fast "System 1" layer for a CAD agent. You give it a goal, an ordered list of features like *"plate 40×30×10 → Ø6 hole at (10, 0) → polar pattern ×6 → fillet the top edges"*. It builds the part command by command: select a plane, sketch, draw, constrain, pad, pattern, fillet. At every step it reads FreeCAD's live state (feature tree, selection, sketch constraints, workbench) and picks the next command from the ones currently available.
+
+- **Tiny and fast.** 1.2M parameters, trained from scratch, ~1 ms per decision on a CPU. No LLM, no vision model, no screenshots.
+- **Generalizes to longer parts.** Trained on parts with at most 5 features, it builds 11-feature parts (~55 commands) with 100% success and 17-feature parts at 95%. The previous version scored 0% at 6+ features.
+- **Recovers from mistakes.** With 20% of its actions replaced by random ones, it notices the damage, undoes it and finishes 86–100% of parts.
+- **Runs in the real FreeCAD app.** It drives the FreeCAD GUI over a local socket and builds parts live.
+
+## Results
+
+![Parts built correctly vs. goal length](assets/length.png)
+
+| Goal | Built correctly | With 20% random actions injected |
+|---|---|---|
+| Parts like the training set (1–5 features) | 100% | 93–100% |
+| 6–7 features | 100% | 86% |
+| 8–9 features | 100% | 87% |
+| 11 features (~55 commands) | 100% | 90% |
+| 13 / 15 / 17 features | 100 / 100 / 95% | – |
+| Feature combinations never seen in training | 90–100% | 94–97% |
+
+"Built correctly" means the model finished and the final solid matches the target exactly (volumetric IoU ≥ 0.99, no stray objects). Each row is 100 fresh goals in FreeCAD 1.1 (60 per length for 13–17 features). Per-step accuracy against the teacher's choices is 99.8%.
+
+## What made it generalize
+
+![What made it generalize](assets/ablation.png)
+
+1. **Randomized position IDs during training** ([Ruoss et al. 2023](https://arxiv.org/abs/2305.16843)). Position numbers the model had never seen were what broke it on longer parts.
+2. **Coupled ordinals.** Goal item *k* and the *k*-th feature in the tree share an index ([position coupling](https://arxiv.org/abs/2405.20671)).
+3. **A modular "done?" policy.** Each goal item asks "am I built yet?", and the model acts on the first one that isn't. This is what made the longest goals reliable across training seeds.
+4. **Factorized feature types.** Category embeddings plus type dropout help with feature pairings it hasn't seen.
+
+## Usage
+
+```python
+from freecad_s1.model.net import from_pretrained
+from freecad_s1.rollout import Policy
+
+model = from_pretrained("shhivv/taiga-s1")
+policy = Policy(model, device="cpu")
+
+probs = policy.score(state, goal, actions)   # {command: probability}, best first
+next_command = next(iter(probs))
+```
+
+What to pass in:
+- `state`: a snapshot of the FreeCAD session, from the included runtime (`freecad_s1.runtime`).
+- `goal`: the ordered feature list, plus a rough size of the finished part (bounding box, volume). Estimates are fine.
+- `actions`: the commands currently available in FreeCAD, as returned by the runtime's `valid_actions()`.
+
+The returned probabilities are calibrated. The fitted temperature is stored in `config.json`.
+
+## How it was trained
+
+- **Data.** 24k synthetic modeling sessions scripted in headless FreeCAD, about 590k decisions. A scripted teacher labels the right next command at every step, and random mistakes are mixed in so the model also learns to recover.
+- **Training.** Supervised training, then two rounds of DAgger: the model drives FreeCAD itself and the teacher corrects what it gets wrong.
+- **Architecture.** A 3-layer transformer encodes the session state and the goal; candidate commands attend to the state and the active goal item, and each gets one score.
+
+**Scope.** It covers FreeCAD PartDesign workflows: sketches (rectangle, circle, hexagon), pad, pocket, hole, revolve, linear and polar patterns, mirror, fillet, chamfer and shell. Taiga-S1 chooses the command; numeric values come from the goal.
+
+## Citation
+
+```bibtex
+@misc{taiga_s1_2026,
+  title  = {Taiga-S1: a small next-action model for CAD},
+  author = {Shanmugam, Shiv},
+  year   = {2026}
+}
+```

config.json

@@ -0,0 +1,51 @@
+{
+  "architecture": "S1Model",
+  "config": {
+    "width": 128,
+    "heads": 4,
+    "enc_layers": 3,
+    "dec_layers": 2,
+    "ff": 384,
+    "dropout": 0.1,
+    "id_dropout": 0.1,
+    "pos_mode": "rand",
+    "pos_table": 48,
+    "ordinal": true,
+    "ord_table": 12,
+    "progress_head": false,
+    "invariant_numerics": true,
+    "modular": true,
+    "pointer": "done",
+    "index_eval": "identity",
+    "type_dropout": 0.15,
+    "temperature": 2.554
+  },
+  "metadata": {
+    "name": "taiga-s1",
+    "version": "0.3.0",
+    "trained_on": "synthetic FreeCAD PartDesign episodes",
+    "calibration": {
+      "states": 11300,
+      "episodes": 256,
+      "T": 2.554,
+      "held_out_T1": {
+        "n": 5573,
+        "acc": 0.9551,
+        "nll": 0.3595,
+        "ece": 0.0416,
+        "mean_conf": 0.9921,
+        "mean_conf_when_wrong": 0.9705,
+        "n_wrong": 250
+      },
+      "held_out_T": {
+        "n": 5573,
+        "acc": 0.9551,
+        "nll": 0.1785,
+        "ece": 0.0272,
+        "mean_conf": 0.9545,
+        "mean_conf_when_wrong": 0.8915,
+        "n_wrong": 250
+      }
+    }
+  }
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ad63f77399eaddd3aebae0d0cf493340885d741a9fe51c1e4737f7273a1408c
+size 4924516