Use standard Hugging Face downloads and local evaluation

README.md

@@ -12,35 +12,36 @@ tags:
 One selected checkpoint per task, with portable benchmark/calibration assets and
 measured results. This repository stays private during testing.
 
-## Install and evaluate
+## Setup
 
-GitHub installation only; nothing has been published to PyPI. A CUDA GPU is recommended.
+GitHub installation only; nothing is published to PyPI. A CUDA GPU is recommended.
 
 ```bash
 python -m venv .venv
 source .venv/bin/activate
 pip install "git+https://github.com/helblazer811/FlowMapReasoning.git@public-simplified"
-hf auth login  # private repository: use an authorized account
-frm-eval --task sudoku
-frm-eval --task maze_unique
-frm-eval --task zebra
+hf auth login  # while this model repository is private
 ```
 
-Or from Python:
+## Download and evaluate
+
+Use standard Hugging Face downloads:
 
 ```python
-from frm import evaluate
-result = evaluate("sudoku", output_dir="outputs/sudoku")
-print(result["policies"][-1])
+from huggingface_hub import snapshot_download
+snapshot_download("helblazer811/FlowReasoningModel", allow_patterns=["sudoku/*"], local_dir="models")
+```
+
+```bash
+python -m frm.evaluate --model-dir models/sudoku
 ```
 
-All model files, benchmark arrays and calibration are downloaded automatically and
-verified by SHA-256. The default is calibrated sequential restarts. Add
-`--max-samples 8` for a functional smoke test, not a reported benchmark score.
-Use `--revision <Hub commit>` to pin the bundle. Outputs resume automatically.
-The evaluator computes a fixed-depth comparison too: evaluation wall-clock work
-includes that extra baseline; reported NFE/FLOPs describe the selected restart
-policy, not the total cost of generating all three comparison rows.
+Replace `sudoku` with `maze_unique` or `zebra` for the other models. Evaluation
+checks the downloaded files and uses calibrated sequential restarts. Results go
+under `outputs/`. Add `--max-samples 8` for a functional check, not a benchmark score.
+The evaluator computes fixed-depth comparisons too; reported policy NFE excludes
+that additional comparison work. For manual download, obtain all files in the
+corresponding model folder from this repository.
 
 ## Selected models and measured results
 
@@ -77,7 +78,7 @@ Each folder contains `model.pt`, `training.yaml`, `threshold.json`, `test.npz`,
 The manifest records the original checkpoint hash and the compact inference-checkpoint
 hash. Optimizer state is omitted; raw weights and EMA are retained. These are compact PyTorch checkpoints from the recorded training runs.
 For training from scratch, download `training.yaml` and run
-`frm-train --config training.yaml --seed <recorded seed>` on the matching data.
+`python -m frm.train --config training.yaml` on the matching data.
 
 Code is being drafted on
 [public-simplified](https://github.com/helblazer811/FlowMapReasoning/tree/public-simplified).