Ocean Blue: Blue-3

Built with Qwen. Named by Thomas. Blue-3 is Qwen2.5-7B-Instruct (the ancestor) fine-tuned in the Ocean Blue loop. A Qwen2.5-72B-Instruct teacher answered code-graded problems, and only answers the grader verified became lessons (803 verified solutions). Blue-3 then attempted harder problems, including 685 the teacher had failed, and trained again on its own verified wins (731 kept, 132 on problems the teacher failed). A self-distillation stage is kept only if it does not lose on the held-out test. Final stage: stage2. A small personality layer (304 replies written by the teacher in Ocean Blue's voice, kept only at 8/10 or above from a judge) gives it its voice. No Claude or other closed-model output was used for training.

Held-out results (code-graded, never trained on)

Model Held-out Deep water (harder)
Ancestor Qwen2.5-7B (untrained) 12.4% 7.0%
Blue-3 after the teacher 50.7% 50.0%
Blue-3 after self-distillation 53.3% 49.0%
Teacher Qwen2.5-72B 36.9% 38.0%

License

Weights: Apache-2.0 (from Qwen2.5-7B-Instruct). Training data was generated by Qwen2.5-72B-Instruct, which is under the Qwen License Agreement (see LICENSE_QWEN_TEACHER); as that license asks: Built with Qwen.

Downloads last month
12
Safetensors
Model size
8B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for cipherxsnier/ocean-blue-blue3

Base model

Qwen/Qwen2.5-7B
Adapter
(2,980)
this model
Adapters
1 model