初始化项目，由ModelHub XC社区提供模型

Model: promotion/qwen3-8b-aaai27-flagship-sppo-avg-s44 Source: Original Platform
2026-07-21 13:43:13 +08:00
commit 9ad339e5b2
21 changed files with 4225 additions and 0 deletions
--- a/pair_manifest.json
+++ b/pair_manifest.json
@@ -0,0 +1,34 @@
+{
+  "schema_version": 1,
+  "objectives": [
+    "helpfulness",
+    "safety",
+    "conciseness"
+  ],
+  "normalization": "per_prompt_minmax",
+  "adversary": {
+    "steps": 25,
+    "alpha": 1.0,
+    "kappa": 0.05,
+    "preference_scale": 8.0
+  },
+  "ronpo_estimator": "Rao-Blackwellized full sigma expectation with three common response pairs per prompt",
+  "counts": {
+    "train": {
+      "avg": 16746,
+      "ronpo_full_expect": 50340,
+      "ronpo_k_only": 50340,
+      "ht_mnpo_helpfulness": 16743,
+      "ht_mnpo_safety": 16764,
+      "ht_mnpo_conciseness": 16755
+    },
+    "validation": {
+      "avg": 1859,
+      "ronpo_full_expect": 5592,
+      "ronpo_k_only": 5592,
+      "ht_mnpo_helpfulness": 1864,
+      "ht_mnpo_safety": 1862,
+      "ht_mnpo_conciseness": 1858
+    }
+  }
+}