Ollaya package for amyrmahdy/decima-small

README.md

@@ -0,0 +1,48 @@
+---
+license: apache-2.0
+base_model:
+- amyrmahdy/decima-small
+- amyrmahdy/decima-base
+- amyrmahdy/decima-agent
+library_name: onnx
+tags:
+- ollaya
+- onnx
+- decision-model
+- system-one
+pipeline_tag: text-classification
+---
+
+# decima for Ollaya
+
+[Ollaya](https://github.com/ollaya-dev/ollaya) package of **[amyrmahdy/decima-small](https://huggingface.co/amyrmahdy/decima-small)** and **[amyrmahdy/decima-base](https://huggingface.co/amyrmahdy/decima-base)** and **[amyrmahdy/decima-agent](https://huggingface.co/amyrmahdy/decima-agent)** by A. M. Madani (amyrmahdy).
+Ollaya runs open decision models locally, the way Ollama runs LLMs: typed questions in,
+calibrated answers out, behind a TypeSafe-compatible API.
+
+```sh
+ollaya run decima
+```
+
+## What is in this repository
+
+This repository holds only the files Ollaya derives, with no weights. Each graph is an ONNX export of the
+original model whose weights **reference the authors' own weight files by byte offset**,
+so `ollaya pull` downloads the weights from the upstream repositories, unmodified and pinned to a
+commit, and verifies their sha256.
+
+| Tag | Upstream | Files |
+|---|---|---|
+| `decima:small` | [amyrmahdy/decima-small@2e7f4d0](https://huggingface.co/amyrmahdy/decima-small/tree/2e7f4d0757df0215f48f2a9b2b589e1f3a6348ed) | `small/model-fp32.onnx`, `small/decision.json`, `small/calibration.json` |
+| `decima:base` | [amyrmahdy/decima-base@2468005](https://huggingface.co/amyrmahdy/decima-base/tree/2468005d5e48e95eb74072c32a6d9df164578071) | `base/model-fp32.onnx`, `base/decision.json`, `base/calibration.json` |
+| `decima:agent` | [amyrmahdy/decima-agent@86a07aa](https://huggingface.co/amyrmahdy/decima-agent/tree/86a07aab1c340fa5869bdb754e57d3851a1d288a) | `agent/model-fp32.onnx`, `agent/decision.json`, `agent/calibration.json` |
+
+Each tag has an fp32 graph, used on CPU and GPU. Each tag also has `decision.json` (sequence layout, special tokens) and
+`calibration.json` (temperatures).
+
+## Parity
+
+Ollaya's Rust runtime matches the author's own code (decima/model.py and systemone.py, fp32) on 581 questions from 122 requests for each of small, base and agent, on CPU and CUDA (RTX 4090 and RTX 5090): identical token rows and truncation, the same 18 rejected requests, the same decision on every question, scores within 2.0e-5 and probabilities within 5.5e-6.
+
+## License
+
+Same as the upstream model (Apache-2.0). Ollaya itself is Apache-2.0.

agent/calibration.json

@@ -0,0 +1,9 @@
+{
+  "temperature": [
+    0.6081955235647175,
+    1.0,
+    0.6081955235647175
+  ],
+  "temperature_by_options": {},
+  "source": "pytorch/decima.json `temperature`, fitted by the author. Score questions take it inside the ordinal head (decision.json `temperature`), so their slot here is 1."
+}

agent/decision.json

@@ -0,0 +1,124 @@
+{
+  "engine": "onnx",
+  "family": "decima",
+  "layout": "decima-late-interaction-v1",
+  "upstream": {
+    "repo": "amyrmahdy/decima-agent",
+    "revision": "86a07aab1c340fa5869bdb754e57d3851a1d288a",
+    "tag": "v2.1",
+    "checkpoint": "pytorch",
+    "code": {
+      "repo": "https://github.com/amyrmahdy/decima",
+      "tag": "v1.1.1",
+      "commit": "2df60942c68b0e1dfc87743462745bacdd936d8a"
+    },
+    "license": "Apache-2.0",
+    "base": "amyrmahdy/decima-base, on jhu-clsp/mmBERT-base (MIT)"
+  },
+  "contract": {
+    "inputs": {
+      "state_ids": {
+        "dtype": "int64",
+        "shape": [
+          "questions",
+          "state_len"
+        ],
+        "note": "each question's state row, right-padded with special_tokens.pad"
+      },
+      "state_mask": {
+        "dtype": "int64",
+        "shape": [
+          "questions",
+          "state_len"
+        ],
+        "note": "1 on tokens"
+      },
+      "option_ids": {
+        "dtype": "int64",
+        "shape": [
+          "options",
+          "option_len"
+        ],
+        "note": "every option row, questions in order, right-padded"
+      },
+      "option_mask": {
+        "dtype": "int64",
+        "shape": [
+          "options",
+          "option_len"
+        ],
+        "note": "1 on tokens"
+      },
+      "option_state": {
+        "dtype": "int64",
+        "shape": [
+          "options"
+        ],
+        "note": "the question (state row) of each option"
+      }
+    },
+    "outputs": {
+      "scores": {
+        "dtype": "float32",
+        "shape": [
+          "options"
+        ],
+        "note": "raw option scores; choice and noul: softmax(scores / temperature)"
+      },
+      "ordinal_g": {
+        "dtype": "float32",
+        "shape": [
+          "options"
+        ],
+        "note": "ord_g(z) per option"
+      },
+      "ordinal_gap": {
+        "dtype": "float32",
+        "shape": [
+          "options"
+        ],
+        "note": "ord_gap(z) per option"
+      }
+    }
+  },
+  "state_prefix": "",
+  "option_prefix": "",
+  "question_in_state": true,
+  "question_in_options": true,
+  "max_state_tokens": 2048,
+  "max_option_tokens": 64,
+  "special_tokens": {
+    "cls": 2,
+    "sep": 1,
+    "pad": 0
+  },
+  "noul": {
+    "options": [
+      "yes",
+      "no"
+    ],
+    "criteria": "\nTrue if: {true}\nFalse if: {false}",
+    "missing": "\u2014"
+  },
+  "limits": {
+    "questions": [
+      1,
+      256
+    ],
+    "choices": [
+      2,
+      255
+    ],
+    "levels": [
+      2,
+      10
+    ]
+  },
+  "temperature": 0.6081955235647175,
+  "option_logits": {
+    "choice": "labels in the caller's order",
+    "score": "levels in order",
+    "noul": "[yes, no]"
+  },
+  "opset": 20
+}

agent/model-fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:713a903f75c20edb4cf2646e3eed54247b3565e4e917cabfe18cf87b64bcac0f
+size 5083416

base/calibration.json

@@ -0,0 +1,9 @@
+{
+  "temperature": [
+    0.7269418400253115,
+    1.0,
+    0.7269418400253115
+  ],
+  "temperature_by_options": {},
+  "source": "pytorch/decima.json `temperature`, fitted by the author. Score questions take it inside the ordinal head (decision.json `temperature`), so their slot here is 1."
+}

base/decision.json

@@ -0,0 +1,124 @@
+{
+  "engine": "onnx",
+  "family": "decima",
+  "layout": "decima-late-interaction-v1",
+  "upstream": {
+    "repo": "amyrmahdy/decima-base",
+    "revision": "2468005d5e48e95eb74072c32a6d9df164578071",
+    "tag": "v2.0",
+    "checkpoint": "pytorch",
+    "code": {
+      "repo": "https://github.com/amyrmahdy/decima",
+      "tag": "v1.1.1",
+      "commit": "2df60942c68b0e1dfc87743462745bacdd936d8a"
+    },
+    "license": "Apache-2.0",
+    "base": "jhu-clsp/mmBERT-base (MIT)"
+  },
+  "contract": {
+    "inputs": {
+      "state_ids": {
+        "dtype": "int64",
+        "shape": [
+          "questions",
+          "state_len"
+        ],
+        "note": "each question's state row, right-padded with special_tokens.pad"
+      },
+      "state_mask": {
+        "dtype": "int64",
+        "shape": [
+          "questions",
+          "state_len"
+        ],
+        "note": "1 on tokens"
+      },
+      "option_ids": {
+        "dtype": "int64",
+        "shape": [
+          "options",
+          "option_len"
+        ],
+        "note": "every option row, questions in order, right-padded"
+      },
+      "option_mask": {
+        "dtype": "int64",
+        "shape": [
+          "options",
+          "option_len"
+        ],
+        "note": "1 on tokens"
+      },
+      "option_state": {
+        "dtype": "int64",
+        "shape": [
+          "options"
+        ],
+        "note": "the question (state row) of each option"
+      }
+    },
+    "outputs": {
+      "scores": {
+        "dtype": "float32",
+        "shape": [
+          "options"
+        ],
+        "note": "raw option scores; choice and noul: softmax(scores / temperature)"
+      },
+      "ordinal_g": {
+        "dtype": "float32",
+        "shape": [
+          "options"
+        ],
+        "note": "ord_g(z) per option"
+      },
+      "ordinal_gap": {
+        "dtype": "float32",
+        "shape": [
+          "options"
+        ],
+        "note": "ord_gap(z) per option"
+      }
+    }
+  },
+  "state_prefix": "",
+  "option_prefix": "",
+  "question_in_state": true,
+  "question_in_options": true,
+  "max_state_tokens": 512,
+  "max_option_tokens": 64,
+  "special_tokens": {
+    "cls": 2,
+    "sep": 1,
+    "pad": 0
+  },
+  "noul": {
+    "options": [
+      "yes",
+      "no"
+    ],
+    "criteria": "\nTrue if: {true}\nFalse if: {false}",
+    "missing": "\u2014"
+  },
+  "limits": {
+    "questions": [
+      1,
+      256
+    ],
+    "choices": [
+      2,
+      255
+    ],
+    "levels": [
+      2,
+      10
+    ]
+  },
+  "temperature": 0.7269418400253115,
+  "option_logits": {
+    "choice": "labels in the caller's order",
+    "score": "levels in order",
+    "noul": "[yes, no]"
+  },
+  "opset": 20
+}

base/model-fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14bda1306d8e85243e73ad9cdc154d92679d9f35f6de6b4d96b305d617798aaf
+size 5083412

small/calibration.json

@@ -0,0 +1,9 @@
+{
+  "temperature": [
+    0.9355568358032639,
+    1.0,
+    0.9355568358032639
+  ],
+  "temperature_by_options": {},
+  "source": "pytorch/decima.json `temperature`, fitted by the author. Score questions take it inside the ordinal head (decision.json `temperature`), so their slot here is 1."
+}

small/decision.json

@@ -0,0 +1,124 @@
+{
+  "engine": "onnx",
+  "family": "decima",
+  "layout": "decima-late-interaction-v1",
+  "upstream": {
+    "repo": "amyrmahdy/decima-small",
+    "revision": "2e7f4d0757df0215f48f2a9b2b589e1f3a6348ed",
+    "tag": "v1.1.1",
+    "checkpoint": "pytorch",
+    "code": {
+      "repo": "https://github.com/amyrmahdy/decima",
+      "tag": "v1.1.1",
+      "commit": "2df60942c68b0e1dfc87743462745bacdd936d8a"
+    },
+    "license": "Apache-2.0",
+    "base": "intfloat/multilingual-e5-small (MIT)"
+  },
+  "contract": {
+    "inputs": {
+      "state_ids": {
+        "dtype": "int64",
+        "shape": [
+          "questions",
+          "state_len"
+        ],
+        "note": "each question's state row, right-padded with special_tokens.pad"
+      },
+      "state_mask": {
+        "dtype": "int64",
+        "shape": [
+          "questions",
+          "state_len"
+        ],
+        "note": "1 on tokens"
+      },
+      "option_ids": {
+        "dtype": "int64",
+        "shape": [
+          "options",
+          "option_len"
+        ],
+        "note": "every option row, questions in order, right-padded"
+      },
+      "option_mask": {
+        "dtype": "int64",
+        "shape": [
+          "options",
+          "option_len"
+        ],
+        "note": "1 on tokens"
+      },
+      "option_state": {
+        "dtype": "int64",
+        "shape": [
+          "options"
+        ],
+        "note": "the question (state row) of each option"
+      }
+    },
+    "outputs": {
+      "scores": {
+        "dtype": "float32",
+        "shape": [
+          "options"
+        ],
+        "note": "raw option scores; choice and noul: softmax(scores / temperature)"
+      },
+      "ordinal_g": {
+        "dtype": "float32",
+        "shape": [
+          "options"
+        ],
+        "note": "ord_g(z) per option"
+      },
+      "ordinal_gap": {
+        "dtype": "float32",
+        "shape": [
+          "options"
+        ],
+        "note": "ord_gap(z) per option"
+      }
+    }
+  },
+  "state_prefix": "query: ",
+  "option_prefix": "passage: ",
+  "question_in_state": true,
+  "question_in_options": true,
+  "max_state_tokens": 512,
+  "max_option_tokens": 64,
+  "special_tokens": {
+    "cls": 0,
+    "sep": 2,
+    "pad": 1
+  },
+  "noul": {
+    "options": [
+      "yes",
+      "no"
+    ],
+    "criteria": "\nTrue if: {true}\nFalse if: {false}",
+    "missing": "\u2014"
+  },
+  "limits": {
+    "questions": [
+      1,
+      256
+    ],
+    "choices": [
+      2,
+      255
+    ],
+    "levels": [
+      2,
+      10
+    ]
+  },
+  "temperature": 0.9355568358032639,
+  "option_logits": {
+    "choice": "labels in the caller's order",
+    "score": "levels in order",
+    "noul": "[yes, no]"
+  },
+  "opset": 20
+}

small/model-fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f83c59768e53fba332d8c51dcdf029893c5cbdf660c0efaba8b747bcaa18250f
+size 3340536