XERON-0.4

XERON-0.4 🎯

XERON-0.4 is the fourth release of the XERON family: a typed-decision (System 1) model forked from XERON-0.2 (itself a fine-tune of convaiinnovations/laya's multilingual checkpoint, backbone jhu-clsp/mmBERT-base, 322M params).

It answers typed questions over a state — choice / score / noul (boolean) — in a single forward pass, returning calibrated probabilities. It never generates text, so it cannot hallucinate and cannot emit malformed schemas.

What changed vs 0.2: 0.4 is a short, stabilized refinement pass on the 0.2 checkpoint, not a bigger-data run. The family's 0.3 attempt (2 epochs on a 62k mixed corpus) came out overconfident — its fitted calibration temperatures blew up to [3.14, 5.41, 5.12] and its soft-probability metrics collapsed. Diagnosis: the RLCD policy-gradient term scales as 1/(2σ²), so annealing σ to 0.1 amplified the gradient ~50× and pushed the logits scale upward; epoch-2 loss also rose (1.075 → 1.289 = overfitting).

0.4 therefore trains 1 epoch with a tamed objective: rl_weight 0.5, σ 0.3→0.2, lr_head 5e-5 (was 1e-4), weight_decay 0.02. That recovers accuracy to a new family best while pulling calibration back most of the way.

Training: 1 epoch · 61,876 sequences · ~2.8 h · 2×T4 (fp16) · post-hoc temperature calibration [2.254, 1.810, 3.188].

📈 Results

JevBench v1.3 (public items only, matched subset)

JevBench's frozen set has 534 decisions; only 231 are public (the judge tier is entirely held out), so these are not directly comparable to the published board ranks. Every system below ran the same 231 items with the official harness (fstandhartinger/jevbench, laya_local adapter); Jev/Laya rows come from the benchmark's own per-task artifact.

XERON-0.2 → XERON-0.4

• XERON-0.4 sets a new family best on overall JevBench accuracy (0.558 vs 0.541) and improves the standard tier, hard tier, hard-tier ECE and typed-decisions accuracy.
• Trade-off: calibration did not fully return to 0.2 levels. If you need the softest, best-calibrated probability distributions, prefer XERON-0.2; if you want the highest decision accuracy, use XERON-0.4.

typed-decisions benchmark

LocalLLaMA/typed-decisions test split, 400 cases / 1,400 decisions:

🚀 사용법

pip install laya

import laya

agent = laya.load("PIXELZX/XERON-0.4")

state = "Policy: refunds require a receipt and purchase within 30 days. A customer bought 12 days ago but has no receipt."
questions = {
    "permitted": {"type": "noul", "instructions": "Under the stated policy, is the requested action permitted?"},
    "urgency":   {"type": "score",  "levels": ["0 — no pressure", "1 — routine", "2 — elevated", "3 — critical"]},
}

print(agent.predict(state, questions)["answers"])

📈 재현

git clone https://github.com/PIXELZX0/XERON && cd XERON
# JevBench 공개 231건 (동일 하네스·어댑터)
git clone --depth 1 https://github.com/fstandhartinger/jevbench /tmp/jevbench
python results/jevbench-public/run_public_jevbench.py PIXELZX/XERON-0.4 XERON-0.4 /tmp/jb04
python results/jevbench-public/score_public_jevbench.py /tmp/jb04 XERON-0.4
# typed-decisions
python scripts/evaluate.py --model PIXELZX/XERON-0.4 --split test --device cuda --output eval.json

⚠️ 한계

• 캘리브레이션이 0.2만큼 좋지 않다 (temperature 2.25/1.81/3.19 vs 0.2의 ~1.0). 0.2·0.4 모두 A100이 아닌 하드웨어(fp16)에서 돌린 영향일 수 있으며, bf16 재학습으로 검증할 가치가 있다.
• 4,096 토큰 학습 — 초장문은 CTX_CAP 확장 후 재학습 필요.
• 선택지 개수 상한: Laya 계열 헤드는 head_max_len=256이라 선택지가 많은 태스크(예: 77/151개 intent)는 옵션 텍스트가 잘려 성능이 급락한다. 학습 데이터에서도 그런 config는 제외했다.
• JevBench hard tier(0.351)와 Jev(0.730)의 격차는 여전히 크다 — hard tier는 장문 정책 문서·모호한 트레이드오프·함정 문항 위주로, 우리 학습 데이터에 그런 저작(authored) 루브릭 데이터가 거의 없다.
• 학습 데이터는 대부분 단일 라벨 코퍼스라 확률 충실도(TVD)에 불리하다. soft-label 비중을 늘리면 개선 여지가 있다.

Built on Laya by Convai Innovations (Apache-2.0) and jhu-clsp/mmBERT-base. Data: Jevify jev-bench (mixed licenses, see its manifest).