aider
Datasets
All datasets matching “aider”finland-metsakeskusaider_polyglottunekv-aider-27b
TuneKV Aider-polyglot 27B arms (Qwen3.8-27B hybrid)
Three arms on aider-polyglot with Qwen/Qwen3.8-27B (enable_thinking=false,
16 FA + 48 GDN layers), served by vLLM 0.28 + tunekv_vllm.hybrid connector:
arm
train
n_tune
python-34
olang 5x5 (pinned)
base
—
—
p@1 5/34 (14.7%), p@2 18/34 (52.9%)
p@1 5/25, p@2 15/25
CE+0.1KL
CE (nll 1.0) + 0.1 reverse-KL, lr 1.5e-4, 3ep/195 steps
384
p@1 6/34 (+1), p@2 15/34 (−3)
p@1 5/25 (±0), p@2 16/25 (+1)
GRPO
per-(ex,try)… See the full description on the dataset page: https://huggingface.co/datasets/Zhongzhu/tunekv-aider-27b.DCAgent2_aider_polyglot_DCAgent2_swesmith-stack-reason_20260127_005703aider_polyglot_a3_rl_DCAgent_inferredbugs_sandboxes_verifier_55_8B_20260526_230125aider_polyglot_GLM_4_7_swesmith_sandboxes_with_tests_oracle_verified_120s_maxepf48469bf