Ollaya package for Qwen/Qwen3Guard-Gen-0.6B

0.6b/calibration.json

@@ -0,0 +1,9 @@
+{
+  "temperature": [
+    1.0,
+    1.0,
+    1.0
+  ],
+  "temperature_by_options": {},
+  "source": "raw (no fitted temperature upstream)"
+}

0.6b/decision.json

@@ -0,0 +1,154 @@
+{
+  "engine": "onnx",
+  "family": "qwen3guard",
+  "layout": "qwen3guard-gen-v1",
+  "upstream": {
+    "repo": "Qwen/Qwen3Guard-Gen-0.6B",
+    "revision": "fada3b2f655b89601929198343c94cd2f64d93cc"
+  },
+  "preset_only": true,
+  "questions": {
+    "safety": {
+      "type": "choice",
+      "instructions": "Qwen3Guard safety level of the user text",
+      "criteria": {
+        "safe": null,
+        "controversial": null,
+        "unsafe": null
+      }
+    },
+    "unsafe": {
+      "type": "noul",
+      "instructions": "Qwen3Guard: the user text is unsafe"
+    },
+    "unsafe_strict": {
+      "type": "noul",
+      "instructions": "Qwen3Guard: the user text is unsafe or controversial"
+    },
+    "category": {
+      "type": "choice",
+      "instructions": "Qwen3Guard: first unsafe category, given unsafe",
+      "criteria": {
+        "none": null,
+        "violent": null,
+        "non_violent_illegal": null,
+        "sexual": null,
+        "pii": null,
+        "suicide_self_harm": null,
+        "unethical": null,
+        "politically_sensitive": null,
+        "copyright": null,
+        "jailbreak": null
+      }
+    }
+  },
+  "contract": {
+    "inputs": {
+      "input_ids": {
+        "dtype": "int64",
+        "shape": [
+          "rows",
+          "seq"
+        ],
+        "note": "right-padded, any pad id"
+      },
+      "last_pos": {
+        "dtype": "int64",
+        "shape": [
+          "rows"
+        ],
+        "note": "index of the last real token"
+      }
+    },
+    "outputs": {
+      "cand_logits": {
+        "dtype": "float32",
+        "shape": [
+          "rows",
+          13
+        ],
+        "note": "[0:3] safety (safe, controversial, unsafe); [3:13] categories"
+      }
+    },
+    "attention": "causal; no mask input",
+    "positions": "0..seq-1, implicit"
+  },
+  "templates": {
+    "prompt_head": "<|im_start|>user\n# Task:\n1) Evaluate whether THE LAST USER's query belongs to safe or unsafe content based on our safety policy.\n2) Identify the category of unsafe content found.\n\n<BEGIN SAFETY POLICY>\n**Safety: Safe**: The USER's query does not contain any prohibited content or content that could lead to prohibited content.\n**Safety: Unsafe**: The USER's query contains prohibited content.\n**Safety: Controversial**: The USER's query does not explicitly contain prohibited content, but its intent, context, or potential responses could be misinterpreted, misused, or exploited to generate unsafe content under certain conditions.\n<END SAFETY POLICY>\n\n<BEGIN UNSAFE CONTENT CATEGORIES>\nViolent.\nNon-violent Illegal Acts.\nSexual Content or Sexual Acts.\nPII.\nSuicide & Self-Harm.\nUnethical Acts.\nPolitically Sensitive Topics.\nCopyright Violation.\nJailbreak.\n<END UNSAFE CONTENT CATEGORIES>\n\n<BEGIN CONVERSATION>\n\nUSER: ",
+    "prompt_tail": "\n\n<END CONVERSATION>\n\nProvide your safety assessment for ONLY THE LAST **USER's query** in the above conversation:\n - The first line must be one of: 'Safety: Safe', 'Safety: Unsafe', 'Safety: Controversial'.\n - The second line should start with 'Categories:' followed by a list of any unsafe content categories, separated by commas. If the content is safe, use 'Categories: None'.<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n",
+    "state": "strings verbatim, else json.dumps(ensure_ascii=False)",
+    "row_safety": "{prompt_head}{state}{prompt_tail}Safety:",
+    "row_category": "{prompt_head}{state}{prompt_tail}Safety: Unsafe\nCategories:",
+    "tokenize": "whole row string, add_special_tokens=False, special tokens parsed"
+  },
+  "candidates": {
+    "safety": {
+      "labels": [
+        "safe",
+        "controversial",
+        "unsafe"
+      ],
+      "texts": [
+        " Safe",
+        " Controversial",
+        " Unsafe"
+      ],
+      "ids": [
+        22291,
+        2093,
+        73067
+      ]
+    },
+    "category": {
+      "labels": [
+        "none",
+        "violent",
+        "non_violent_illegal",
+        "sexual",
+        "pii",
+        "suicide_self_harm",
+        "unethical",
+        "politically_sensitive",
+        "copyright",
+        "jailbreak"
+      ],
+      "texts": [
+        " None",
+        " Violent",
+        " Non-violent Illegal Acts",
+        " Sexual Content or Sexual Acts",
+        " PII",
+        " Suicide & Self-Harm",
+        " Unethical Acts",
+        " Politically Sensitive Topics",
+        " Copyright Violation",
+        " Jailbreak"
+      ],
+      "ids": [
+        2240,
+        29455,
+        11581,
+        38667,
+        393,
+        64800,
+        1230,
+        15928,
+        2955,
+        56146
+      ]
+    }
+  },
+  "option_logits": {
+    "safety": "cand[r0, 0:3]",
+    "unsafe": "[logsumexp(cand[r0,0], cand[r0,1]), cand[r0,2]]",
+    "unsafe_strict": "[cand[r0,0], logsumexp(cand[r0,1], cand[r0,2])]",
+    "category": "cand[r1, 3:13]"
+  },
+  "max_len": 32768,
+  "special_tokens": {
+    "pad": 151643,
+    "bos": null,
+    "add_special_tokens": false
+  },
+  "opset": 20
+}

0.6b/model-fp32.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e0f89fc275e2bb0d459248abaa0d97ed50f360e4e3a5f0c5b1b29af0765d312
+size 5426737

0.6b/questions.json

@@ -0,0 +1,35 @@
+{
+  "safety": {
+    "type": "choice",
+    "instructions": "Qwen3Guard safety level of the user text",
+    "criteria": {
+      "safe": null,
+      "controversial": null,
+      "unsafe": null
+    }
+  },
+  "unsafe": {
+    "type": "noul",
+    "instructions": "Qwen3Guard: the user text is unsafe"
+  },
+  "unsafe_strict": {
+    "type": "noul",
+    "instructions": "Qwen3Guard: the user text is unsafe or controversial"
+  },
+  "category": {
+    "type": "choice",
+    "instructions": "Qwen3Guard: first unsafe category, given unsafe",
+    "criteria": {
+      "none": null,
+      "violent": null,
+      "non_violent_illegal": null,
+      "sexual": null,
+      "pii": null,
+      "suicide_self_harm": null,
+      "unethical": null,
+      "politically_sensitive": null,
+      "copyright": null,
+      "jailbreak": null
+    }
+  }
+}

README.md

@@ -0,0 +1,44 @@
+---
+license: apache-2.0
+base_model:
+- Qwen/Qwen3Guard-Gen-0.6B
+library_name: onnx
+tags:
+- ollaya
+- onnx
+- decision-model
+- system-one
+pipeline_tag: text-classification
+---
+
+# qwen3guard for Ollaya
+
+[Ollaya](https://github.com/ollaya-dev/ollaya) package of **[Qwen/Qwen3Guard-Gen-0.6B](https://huggingface.co/Qwen/Qwen3Guard-Gen-0.6B)** by the Qwen team, Alibaba Cloud.
+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 qwen3guard
+```
+
+## 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 |
+|---|---|---|
+| `qwen3guard:0.6b` | [Qwen/Qwen3Guard-Gen-0.6B@fada3b2](https://huggingface.co/Qwen/Qwen3Guard-Gen-0.6B/tree/fada3b2f655b89601929198343c94cd2f64d93cc) | `0.6b/model-fp32.onnx`, `0.6b/decision.json`, `0.6b/calibration.json`, `0.6b/questions.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). `questions.json` holds the built-in questions: the model answers those and no others, so requests leave `questions` out.
+
+## Parity
+
+Ollaya's Rust runtime matches the transformers fp32 reference exactly on 74 texts (148 rows, 296 built-in questions). The token ids are identical, and so is the decision on every question. Probabilities are within 1.4e-5, on CPU and CUDA.
+
+## License
+
+Same as the upstream model (Apache-2.0). Ollaya itself is Apache-2.0.