Upload EPIC router checkpoints

0.25/router_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2f3dc4e1573546449f2ba2181a3364735d33f47474290d63e8fdc7c38a4034a
+size 91514922

0.5/router_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7cae3a9ce9bf889763b5c096f87f97eca80eea53bc9f160a4568d0ff78d60a23
+size 91514922

0.75/router_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0d609bb7a40b9ff7fe262c79fa6f6fd2123a8b30ee50f80dfd3465655ce3ea0
+size 91514922

1.0/router_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ce1209e508d31c52d9f1a7a2d86a8bb49cbf52e502e8011d657044a74eb243b
+size 91514922

README.md

@@ -0,0 +1,80 @@
+# EPIC Router Family
+
+This repository hosts the public checkpoints for the EPIC router models. Each
+checkpoint learns to pick the best reasoning configuration (method, aggregator,
+sample count, etc.) given a natural-language math question. The underlying
+training and evaluation code lives in the EPIC GitHub project; these weights are
+ready-to-use drop-in artifacts for that codebase.
+
+
+## Available Versions
+
+| Subdirectory | File              | Notes                               |
+|--------------|-------------------|-------------------------------------|
+| `0.25/`      | `router_model.pt` | Router trained for the cost-accuracy trade-off = 0.25. |
+| `0.5/`       | `router_model.pt` | Router trained for the cost-accuracy trade-off = 0.5. |
+| `0.75/`      | `router_model.pt` | Router trained for the cost-accuracy trade-off = 0.75 |
+| `1.0/`       | `router_model.pt` | Router trained for the cost-accuracy trade-off = 1.0 |
+
+Each checkpoint contains:
+
+- `state_dict`: PyTorch weights for `RouterScoringModel`
+- `model_name`: base encoder identifier (defaults to `sentence-transformers/all-MiniLM-L6-v2`)
+- `projection_dim`: dimension of the projection head
+- `methods`: serialized reasoning configurations; each row corresponds to one column in the router head
+
+
+## Quickstart (Python)
+
+Install dependencies:
+
+```bash
+pip install huggingface_hub torch sentence-transformers
+```
+
+Load a checkpoint and route a question:
+
+```python
+from huggingface_hub import hf_hub_download
+import torch
+from router.models import RouterScoringModel, MiniLMQuestionEncoder, QuestionProjector
+from data_schemas.reasoning import ReasoningConfig
+
+REPO_ID = "baonn/epic"
+VERSION = "1.0"  # or 0.5 / 0.75 / 0.25
+
+checkpoint_path = hf_hub_download(
+    repo_id=REPO_ID,
+    filename="router_model.pt",
+    subfolder=VERSION,
+)
+checkpoint = torch.load(checkpoint_path, map_location="cpu")
+
+encoder = MiniLMQuestionEncoder(
+    model_name=checkpoint.get("model_name", "sentence-transformers/all-MiniLM-L6-v2"),
+    trainable=False,
+)
+projector = QuestionProjector(
+    input_dim=encoder.transformer.config.hidden_size,
+    projection_dim=int(checkpoint["projection_dim"]),
+)
+model = RouterScoringModel(
+    question_encoder=encoder,
+    projector=projector,
+    num_methods=len(checkpoint["methods"]),
+)
+model.load_state_dict(checkpoint["state_dict"])
+model.eval()
+
+reasoning_configs = [
+    ReasoningConfig.deserialize(payload) for payload in checkpoint["methods"]
+]
+
+question = "How many positive divisors does 3600 have?"
+with torch.no_grad():
+    logits = model([question])
+    method_index = torch.argmax(logits, dim=1).item()
+
+selected_config = reasoning_configs[method_index]
+print("Recommended config:", selected_config.serialize(include_samples=True))
+```