Issues

• RemoteRouter's default Python-urllib User-Agent is blocked by Cloudflare (error 1010) in front of hosted laya-serve deployments

  NandhaKishorM/laya#1361 · Filter by author liyikradlee opened on Oct 25, 2026
  2

• [feature]: local predictLong for the TypeScript Agent and Router

  NandhaKishorM/laya#1316 · Filter by author xinpingzh opened on Oct 22, 2026
  3

• [feature]: YOLO-style tiered checkpoints (Nano/Base/Large) for varying decision complexity

  enhancement
  NandhaKishorM/laya#1303 · Filter by author fly2rain opened on Oct 21, 2026
  1

• [feature]: paired context-robustness evaluation for choice decisions

  enhancement
  NandhaKishorM/laya#1266 · Filter by author xinpingzh opened on Oct 17, 2026
  1

• [bug]: the published English and typed-decisions checkpoints ship choice:11+ = 0.1006, outside the runtime clamp

  NandhaKishorM/laya#1245 · Filter by author gopaljigaur opened on Oct 15, 2026
  5

• [bug]: a noul question containing "positive" answers 0.000 at confidence 1.0 for any state

  NandhaKishorM/laya#1244 · Filter by author gopaljigaur opened on Oct 15, 2026
  6

• [bug]: Discussion: Robustness of Structured Decisions Under Untrusted Context

  bug
  NandhaKishorM/laya#1229 · Filter by author xiaoxiaohg8d opened on Oct 14, 2026
  1

• [bug]: LayaActive/Passive Voice Classification Unsatsifactory

  bug
  NandhaKishorM/laya#1197 · Filter by author QuantumNovice opened on Oct 13, 2026
  1

• [feature]: packed fine-tuning in laya.train, ≈2.7× faster epochs, checkpoints stay standard (follow-up to #49)

  NandhaKishorM/laya#1070 · Filter by author sathwikbs opened on Oct 10, 2026
  3

• False-alarm rise on benign / Spanish inputs since v0.4.0 (v0.3.29 -> v0.4.0/v0.4.1, recall unchanged)

  NandhaKishorM/laya#1058 · Filter by author yemudo opened on Oct 8, 2026
  2

• laya.train: the fixed 4-epoch default collapses to the class prior on small datasets (measured; proposal: early stopping on the calibration slice + a collapse warning)

  NandhaKishorM/laya#963 · Filter by author JaviMaligno opened on Oct 5, 2026
  2

• [feature]: serve your own checkpoints beside the built-ins, routed by what the request asks

  enhancement
  NandhaKishorM/laya#919 · Filter by author alandefreitas opened on Oct 4, 2026
  10

• [bug]: guard: prompt_injection scores 0.00 to 0.30 on instructions embedded in document-shaped text (3 cases, with a matched control)

  bug
  NandhaKishorM/laya#781 · Filter by author rpriven opened on Sep 30, 2026
  5

• [bug]: Probabilities should be independent for each criterion

  bug
  NandhaKishorM/laya#779 · Filter by author smoyer64 opened on Sep 30, 2026
  3

• Independent reproduction: all four published benchmark claims reproduce within 0.0003

  NandhaKishorM/laya#776 · Filter by author Gjusev opened on Sep 30, 2026
  1

• Controlled A/B: GRPO + proper-reward term shows no measurable gain over (soft) cross-entropy in the fine-tune loop

  NandhaKishorM/laya#741 · Filter by author PengyiZhang opened on Sep 29, 2026
  1

• [feature]: adding datasets

  enhancement
  NandhaKishorM/laya#709 · Filter by author Teglgaard opened on Sep 29, 2026
  3

• [feature]: Proposal: Persian (fa) diagnostic benchmark, in the style of feishu_zh

  enhancement
  NandhaKishorM/laya#703 · Filter by author alipyth opened on Sep 28, 2026
  1

• [feature]: Stability under option reorder or rename as an opt-in reliability signal, alongside min_confidence

  enhancement
  NandhaKishorM/laya#635 · Filter by author Charanraj-24 opened on Sep 27, 2026
  8

• [bug]: is_coding scores a calculus prompt at 0.88

  NandhaKishorM/laya#622 · Filter by author jamesonBradfield opened on Sep 27, 2026
  2

• Independent eval: 1,240 scored decisions over 9 suites, laya 0.686 vs Jev 0.907 vs Qwen2.5-1.5B PCD 0.605, with paired intervals

  NandhaKishorM/laya#555 · Filter by author instax-dutta opened on Sep 26, 2026
  1

• Independent eval: 12,000 US federal IT solicitations, 741 quote-graded rows: laya 0.780 vs Jev 0.919, and a 16-point drop at 150 to 300 tokens with the state uncut

  NandhaKishorM/laya#450 · Filter by author bhushankinge opened on Sep 24, 2026
  5

• CUDA default autocast (bf16) is the least precise dtype: up to 0.073 from fp32 and a preset answer flips; fp16 is within 0.019 at the same latency

  NandhaKishorM/laya#443 · Filter by author bhushankinge opened on Sep 24, 2026
  5

• CI: a required check that never runs, cancelled runs on main, and a plan for the rest

  NandhaKishorM/laya#399 · Filter by author maxfield-allison opened on Sep 24, 2026
  10

• Docs site: structure and the guides still to write

  NandhaKishorM/laya#390 · Filter by author maxfield-allison opened on Sep 24, 2026
  21