CoolFace
Datasetpublic

Zhongzhu/tunekv-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.

sourceHugging Faceupdated 4d agoView on Hugging Face
0likes403downloads
Dataset Card

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:

armtrainn_tunepython-34olang 5x5 (pinned)
base——p@1 5/34 (14.7%), p@2 18/34 (52.9%)p@1 5/25, p@2 15/25
CE+0.1KLCE (nll 1.0) + 0.1 reverse-KL, lr 1.5e-4, 3ep/195 steps384p@1 6/34 (+1), p@2 15/34 (−3)p@1 5/25 (±0), p@2 16/25 (+1)
GRPOper-(ex,try) z-scored outcome RL, clip 0.2 + 0.1 reverse-KL, lr 1.5e-4, 2ep/938 steps384p@1 4/34 (−1), p@2 13/34 (−5)p@1 2/25 (−3), p@2 15/25 (±0)

Prefix: LCP 411 tokens of live aider prompts → n_tune=384 (÷64 GDN boundary), capture-init. Eval: whole-format, 2 tries, temp 0, threads 4. Olang = 5 pinned exercises/language (keywords-pinned; --num-tests is an unseeded shuffle).

Training data

  • —25 rollout seeds (10×T0.7 + 4×greedy + 2×T1.0 + 3×T1.3 + 6×hard-full; 282 exercises/seed, full 6-language polyglot).
  • —CE: 395 selected → 373 rows (317 try1 + 56 try2; py94/ja102/js104/cpp73).
  • —GRPO: census 459 informative cells → 2029 rows (py323/ja563/js466/cpp677).
  • —go/rust contribute zero training rows: rollout-container infra failure (go.mod from exercism-HEAD requires go1.26 > container go1.21, offline; RUSTUP_HOME=/opt/rustup override points at an empty dir). Eval (olang) runs on polyglot-benchmark (go1.18) and covers all 6 languages.

Layout

  • —artifacts/ce/, artifacts/grpo/: connector-ready TP1 artifacts (artifact/ + EXPORT_AUDIT.json + PASS), n_inject=384.
  • —rows/: training rows + prefix ids + select/census/build reports.
  • —eval/: per-run results.json/behavior.json/logs (base/CE python-34, base/CE olang per-language; GRPO pending).
  • —rollout_logs/: 25 proxy logs (full request/response bodies) + schedule.
  • —scripts/: row builders, census, train/export yamls, eval drivers.

Provenance

Built 2026-09-22/23 (UTC) on charlie H100 (zz-aider27-20260922): rollout+serve on GPUs 0-1, training on GPUs 2-3. Row builders replicate the 30B-V4 recipe cell-for-cell (census) and value-for-value (sample row). Connector injection verified per-arm (injected 384-token prefix fa16 gdn48).