README.md
| 1 | --- |
| 2 | license: apache-2.0 |
| 3 | library_name: synthefy-nori |
| 4 | pipeline_tag: tabular-regression |
| 5 | tags: |
| 6 | - tabular |
| 7 | - tabular-regression |
| 8 | - tabular-foundation-model |
| 9 | - in-context-learning |
| 10 | - synthetic-data |
| 11 | - pytorch |
| 12 | --- |
| 13 | |
| 14 | <p align="center"> |
| 15 | <img src="synthefy_nori_banner.png" alt="Nori" width="100%"> |
| 16 | </p> |
| 17 | |
| 18 | # Nori-30M |
| 19 | |
| 20 | **Nori-30M** is the ~29.2M-parameter variant of [Nori](https://huggingface.co/Synthefy/Nori), |
| 21 | a tabular foundation model for **regression** via in-context learning (ICL). Given a few labeled |
| 22 | rows as context, it predicts on new query rows in a **single forward pass**, with no task-specific |
| 23 | training or fine-tuning. The model is trained **entirely on synthetic data**. |
| 24 | |
| 25 | - **Documentation:** https://docs.synthefy.com/nori/ |
| 26 | - **Repository:** https://github.com/Synthefy/synthefy-nori |
| 27 | - **Library:** `pip install synthefy-nori` |
| 28 | - **Checkpoint:** `nori.pt` (this repo) |
| 29 | - **Parameters:** ~29.2M |
| 30 | - **License:** Apache-2.0 |
| 31 | |
| 32 | ## Results |
| 33 | |
| 34 | Mean and median R² across 96 regression tasks from three public benchmark suites, on the same |
| 35 | protocol as the [base Nori](https://huggingface.co/Synthefy/Nori): |
| 36 | |
| 37 | | Suite | Datasets | Mean R² | Median R² | |
| 38 | |-------|---------:|--------:|----------:| |
| 39 | | TabArena | 13 | 0.8148 | 0.8834 | |
| 40 | | TALENT | 72 | 0.7575 | 0.8844 | |
| 41 | | OpenML | 11 | 0.6459 | 0.6212 | |
| 42 | | **Overall** | **96** | **0.7525** | **0.8745** | |
| 43 | |
| 44 | Stronger than the ~6M base on every suite. Evaluated with the bundled default inference config |
| 45 | and the large-GPU protocol (up to 50k context rows per dataset). |
| 46 | |
| 47 | ## Use it from your AI coding assistant |
| 48 | |
| 49 | Paste this into Claude Code, Cursor, or any AI coding assistant and it will wire |
| 50 | Nori into your own project: |
| 51 | |
| 52 | ````text |
| 53 | Look at my code/task/report here and figure out where Nori would best fit — it's |
| 54 | Synthefy's tabular foundation model, a drop-in scikit-learn estimator that predicts |
| 55 | a continuous target by in-context learning: no training loop, no hyperparameters, |
| 56 | and it uses the GPU automatically when one's available (CPU otherwise). |
| 57 | |
| 58 | 1. Install it with this project's package manager |
| 59 | (e.g. `uv add synthefy-nori`, or `pip install -U synthefy-nori`). |
| 60 | |
| 61 | 2. Use it wherever a tabular regression / prediction step fits: |
| 62 | |
| 63 | ```python |
| 64 | from synthefy_nori import NoriRegressor |
| 65 | |
| 66 | reg = NoriRegressor(model="nori-30m") # downloads these weights from the Hub on first predict |
| 67 | reg.fit(X_train, y_train) # stores your rows as context — no training happens |
| 68 | y_pred = reg.predict(X_test) # point predictions (predictive-distribution mean) |
| 69 | |
| 70 | # Prediction intervals come free — no conformal/quantile add-ons: |
| 71 | lo, mid, hi = reg.predict(X_test, output_type="quantiles", quantiles=[0.1, 0.5, 0.9]) |
| 72 | ``` |
| 73 | |
| 74 | X is a numeric feature matrix (encode categoricals as ordinals/one-hot, leave |
| 75 | missing values as NaN, no scaling needed); y is a finite continuous target. If I |
| 76 | already have a model, wire Nori up alongside it on the same train/test split and |
| 77 | metric so I can compare them. If the best place to plug Nori in isn't obvious, |
| 78 | show me where you'd put it and confirm with me before making changes. |
| 79 | |
| 80 | Going deeper: synthefy-nori ships a ready-made nori-regression skill for AI coding |
| 81 | assistants with vetted recipes — calibrated prediction intervals, honest baseline |
| 82 | comparison under fixed CV, SHAP/PDP interpretability, and leak-safe one-step |
| 83 | time-series forecasting. Read and follow it if relevant: |
| 84 | https://github.com/Synthefy/synthefy-nori/tree/main/.claude/skills/nori-regression |
| 85 | ```` |
| 86 | |
| 87 | ## Usage |
| 88 | |
| 89 | ```bash |
| 90 | pip install synthefy-nori |
| 91 | ``` |
| 92 | |
| 93 | ```python |
| 94 | from sklearn.datasets import load_diabetes |
| 95 | from sklearn.model_selection import train_test_split |
| 96 | from synthefy_nori import NoriRegressor |
| 97 | |
| 98 | X, y = load_diabetes(return_X_y=True) |
| 99 | X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0) |
| 100 | |
| 101 | model = NoriRegressor(model="nori-30m") # downloads these weights from the Hub on first use |
| 102 | model.fit(X_train, y_train) # "fit" just stores the labeled rows as context |
| 103 | pred = model.predict(X_test) # predictions in a single forward pass, no training |
| 104 | ``` |
| 105 | |
| 106 | It uses a GPU when one is available and falls back to CPU. A one-shot helper skips the object |
| 107 | entirely: |
| 108 | |
| 109 | ```python |
| 110 | from synthefy_nori import predict |
| 111 | pred = predict(X_train, y_train, X_test, task="regression", model="nori-30m") |
| 112 | ``` |
| 113 | |
| 114 | `predict` follows the `TabPFNRegressor.predict` contract: pass `output_type="mean"` (default), |
| 115 | `"median"`, or `"mode"` to choose the point estimate drawn from the model's predictive |
| 116 | distribution. |
| 117 | |
| 118 | To run from a local checkpoint instead of the Hub, pass a path: |
| 119 | `NoriRegressor(model_path="path/to/nori.pt")`. |
| 120 | |
| 121 | This model is **public**: the first call downloads and caches it automatically, with no token and no |
| 122 | access request. A Hugging Face token (read scope) is only worth setting if you hit anonymous |
| 123 | download rate limits — provide it via `export HF_TOKEN=hf_...`, `hf auth login`, or |
| 124 | `NoriRegressor(model="nori-30m", token="hf_...")`. |
| 125 | |
| 126 | ## Intended use & limitations |
| 127 | |
| 128 | - **Intended for** small-to-medium tabular regression where in-context learning is attractive |
| 129 | (no per-task training). |
| 130 | - **Limitations:** dense O(N²) sample attention bounds practical context size, so the current gap |
| 131 | vs the best baselines is on **large-N / long-context** tables. Trained entirely on synthetic |
| 132 | data; no benchmark data is used in training. |
| 133 | |
| 134 | ## Citation |
| 135 | |
| 136 | ```bibtex |
| 137 | @software{synthefy_nori_2026, |
| 138 | title = {Nori: A Tabular Foundation Model Trained on Synthetic Data}, |
| 139 | author = {Synthefy}, |
| 140 | year = {2026}, |
| 141 | url = {https://github.com/Synthefy/synthefy-nori} |
| 142 | } |
| 143 | ``` |
| 144 | |
| 145 | ## License |
| 146 | |
| 147 | Apache-2.0. See |
| 148 | [LICENSE](https://github.com/Synthefy/synthefy-nori/blob/main/LICENSE) and |
| 149 | [NOTICE](https://github.com/Synthefy/synthefy-nori/blob/main/NOTICE). |
| 150 | |