初始化项目，由ModelHub XC社区提供模型

Model: NotHereNorThere/Coral-v1.5-0.6B-raw
Source: Original Platform

mergekit_config.yml

@@ -0,0 +1,43 @@
+# Coral-v1.5-0.6B merge config (mergekit)
+# Method: TIES — computes each finetune's task vector vs the base, trims to the
+# strongest deltas, elects a consensus sign per parameter, then merges. Stable
+# and quality-preserving (same method as CoralLM-1B). Stays at 0.6B params.
+#
+# Run:
+#   mergekit-yaml coral-v1.5-0.6b.yml ./coral-v1.5-0.6b-out --cuda
+#
+# All donors are Qwen3-0.6B finetunes -> shared tokenizer/arch, native <think>.
+
+base_model: Qwen/Qwen3-0.6B          # reference anchor, NOT a contributor
+merge_method: ties
+dtype: bfloat16
+parameters:
+  normalize: true                    # keep merged magnitudes sane across 5 contributors
+  int8_mask: true                    # memory-light sign mask
+
+models:
+  # --- Thinking / reasoning (the spicy stars, split across two sources) ---
+  - model: reaperdoesntknow/Qwen3-0.6B-Distilled-30B-A3B-Thinking-SFT
+    parameters:
+      weight: 0.30
+      density: 0.5
+  - model: MihaiPopa-1/Qwen-3-0.6B-Claude-4.7-Opus-Distilled
+    parameters:
+      weight: 0.30
+      density: 0.5
+
+  # --- Code ---
+  - model: suayptalha/Qwen3-0.6B-Code-Expert
+    parameters:
+      weight: 0.25
+      density: 0.5
+
+  # --- Uncensored / de-alignment (two sources, kept light so they don't dominate) ---
+  - model: DavidAU/Qwen3-0.6B-heretic-abliterated-uncensored
+    parameters:
+      weight: 0.15
+      density: 0.5
+  - model: huihui-ai/Huihui-Qwen3-0.6B-abliterated-v2
+    parameters:
+      weight: 0.15
+      density: 0.5