Ollaya package for PostHog/jeeves

9b/calibration.json

@@ -0,0 +1,9 @@
+{
+  "temperature": [
+    1.8589280843734741,
+    1.8589280843734741,
+    1.8589280843734741
+  ],
+  "temperature_by_options": {},
+  "source": "export.json temperature (one softmax temperature fitted on the dev set)"
+}

9b/decision.json

@@ -0,0 +1,92 @@
+{
+  "engine": "onnx",
+  "family": "jeeves",
+  "layout": "jeeves-markers-v1",
+  "upstream": {
+    "repo": "PostHog/jeeves",
+    "revision": "8622b7d1652a9dcb8629486b84dce9e8d690c5cd",
+    "code": {
+      "repo": "https://github.com/PostHog/jeeves",
+      "commit": "6151619c14fcffb2406830f09fd4f09fdd22200e"
+    },
+    "format": "markers-v3-plainchains",
+    "tokenizer": {
+      "repo": "jaredpalmer/kev-9b",
+      "revision": "2629c06a5aeb0feb3b9783bafed17ed8f39ecf5c",
+      "path": "tokenizer.json"
+    }
+  },
+  "contract": {
+    "inputs": {
+      "input_ids": {
+        "dtype": "int64",
+        "shape": [
+          "rows",
+          "seq"
+        ],
+        "note": "one row per question; seq a multiple of 64; right-pad with the pad id"
+      },
+      "decide_pos": {
+        "dtype": "int64",
+        "shape": [
+          "rows"
+        ],
+        "note": "the row's last position"
+      },
+      "opt_pos": {
+        "dtype": "int64",
+        "shape": [
+          "rows",
+          "k"
+        ],
+        "note": "option-close positions after </think>; shorter rows pad with 0"
+      }
+    },
+    "outputs": {
+      "scores": {
+        "dtype": "float32",
+        "shape": [
+          "rows",
+          "k"
+        ],
+        "note": "raw pointer scores"
+      }
+    },
+    "seq_multiple": 64,
+    "positions": "0..seq-1, implicit",
+    "attention": "causal; no mask input"
+  },
+  "template": {
+    "pre": "<|im_start|>user\n",
+    "post": "<|im_end|>\n<|im_start|>assistant\n<think>\n"
+  },
+  "markers": {
+    "state": "<|fim_prefix|>",
+    "question": "<|fim_middle|>",
+    "option_open": "<|box_start|>",
+    "option_close": "<|box_end|>",
+    "decide": "<|fim_suffix|>",
+    "think_end": "</think>"
+  },
+  "marker_ids": {
+    "state": 248060,
+    "question": 248061,
+    "option_open": 248049,
+    "option_close": 248050,
+    "decide": 248062,
+    "think_end": 248069
+  },
+  "empty_think": "\n",
+  "max_row_tokens": 8192,
+  "pad": 248044,
+  "min_options": 1,
+  "max_options": 255,
+  "option_logits": {
+    "choice": "scores[row, :k]",
+    "noul": "scores[row, :2] (0 = no = false, 1 = yes = true)",
+    "score": "scores[row, :levels]"
+  },
+  "opset": 20,
+  "precision": "fp32 compute; weights BF16 (widened by Cast); head F32",
+  "weights_in_memory": "bf16"
+}

9b/model-fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d171892404203e8c6a5a39a1aa51b38880fa0b19083ebf431bd50e6faaf31038
+size 11093224

README.md

@@ -0,0 +1,44 @@
+---
+license: apache-2.0
+base_model:
+- PostHog/jeeves
+library_name: onnx
+tags:
+- ollaya
+- onnx
+- decision-model
+- system-one
+pipeline_tag: text-classification
+---
+
+# jeeves for Ollaya
+
+[Ollaya](https://github.com/ollaya-dev/ollaya) package of **[PostHog/jeeves](https://huggingface.co/PostHog/jeeves)** by PostHog (fused weights and pointer head) and the Qwen team (base model).
+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 jeeves
+```
+
+## 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 |
+|---|---|---|
+| `jeeves:9b` | [PostHog/jeeves@8622b7d](https://huggingface.co/PostHog/jeeves/tree/8622b7d1652a9dcb8629486b84dce9e8d690c5cd) | `9b/model-fp32.onnx`, `9b/decision.json`, `9b/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 authors' own code (their Qwen3.5 model and pointer head, fp32, no thinking) on 430 questions from 107 requests, on CUDA: identical token rows and option positions, the same 16 rejected requests, the same decision on every question, scores within 1.9e-4 and probabilities within 1.4e-5.
+
+## License
+
+Same as the upstream model (Apache-2.0). Ollaya itself is Apache-2.0.