lkjiop8/Yuanl-27B-v59-long

Yuanl-27B v59-long (LoRA adapter)

LoRA adapter (r=64, alpha=128) in the Yuanl-27B cybersecurity/agentic lineage.

• Architecture base: Qwen3.6-27B (qwen3_5, with native MTP nextn heads in the full upstream checkpoint).
• Lineage: trained on the SFT-merged base of the v5-x → v5-9-sft chain (merged_v6_1_base), then long-context (128K) continue-train. Held-out eval 7/12; with a task-completion-discipline system prompt it reaches 9/12.
• Targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj (language-model layers only; never touches nextn.* / visual.*).

Note: this adapter's true training base is the private SFT-merged merged_v6_1_base, not raw unsloth/Qwen3.6-27B. The base_model field above points at the public architecture base for reference; applying this adapter directly on the raw base reproduces only the final delta, not the full v5-x SFT chain.