slayer-v48-qwen3.5-27b

LoRA adapter for Qwen/Qwen3.5-27B trained for Polish Open PL style tasks.

Status

This adapter should be treated as a research checkpoint unless the attached evaluation artifacts show it was explicitly promoted.

• Local adapter directory: /home/ubuntu/runs/qwen35_27b_sft_openpl_style_v48_full_v43_epoch_lr2e-6_steps14832_seed48027_lora
• Training data: not recorded
• Local Open PL proxy score: n/a
• Promotion decision: not attached
• Bielik published target used for orientation: 65.93

The Bielik number is a published leaderboard target, not a same-harness local Bielik run, unless a separate same-harness Bielik artifact is attached.

Evaluation Artifacts

The repository may include:

• eval/result_summary.json
• eval/open_pl_average.json
• eval/compare_vs_v31_broad_closed200.json
• eval/compare_vs_v48_broad_closed200.json
• eval/compare_vs_bielik_cached_broad_closed200.json
• eval/promotion_decision.json

These are local proxy artifacts. Do not report them as official leaderboard results without a full official evaluation.

Notes

v48 Qwen3.5-27B LoRA full v43-epoch SFT. Best slayer to date; closed non-RAG present avg 69.91 (>v31 63.29, >Bielik target 65.93).

Compare vs v31

{
  "improved": 28,
  "metrics": 38,
  "missing": 0,
  "regressed": 8,
  "unchanged": 2
}