Ocean Blue: Blue-2

Built with Qwen. Blue-2 is Qwen2.5-3B-Instruct fine-tuned (LoRA, merged) inside the Ocean Blue loop: an open teacher (Qwen2.5-14B-Instruct) is rejection-sampled against code-graded puzzles, and only answers the grader verifies are kept as training data (1,464 verified solutions, 3 epochs). No Claude or other closed-model output was used.

Results on a held-out, code-graded test (300 tasks)

Family Blue-2 before Blue-2 after Teacher (14B)
Ledger checking 37% 95% 82%
Log-based claims 60% 100% 92%
Stale vs fresh notes 28% 70% 62%
Line ordering 13% 25% 33%
Code reading 3% 22% 17%
Overall 28% 62.3% 57%

Blue-2 is a specialist for these structured-reasoning families, not a general chat model.

License

Derivative of Qwen2.5-3B-Instruct, distributed under the Qwen RESEARCH LICENSE AGREEMENT (see LICENSE and NOTICE). Non-commercial use only (research or evaluation).

Downloads last month
30
Safetensors
Model size
3B 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-blue2

Base model

Qwen/Qwen2.5-3B
Adapter
(1,524)
this model