A fine-tuned local LLM-as-judge for scoring the psychological safety of conversational-AI transcripts — the offline default judge of aipsy-bench.

aipsy-judge-1.0 reads the full conversation — both the user's and the assistant's turns — and scores only the assistant's responses, judged in the context of the whole conversation arc, on a clinical-safety rubric (empathy, tone, crisis handling, advice safety, boundary safety, affective complexity, + a crisis-detection flag). It runs fully local via Ollama — no API keys, nothing sent to third-party providers — so the entire transcript, including sensitive user content, never leaves your machine.

It is directional, not authoritative. A recommendation, not a rubber-stamp — a false-positive-leaning, human-in-the-loop screen that flags likely-problematic turns for review. It is built to extend a reviewer's reach, not replace them: it is not a safety certifier and not a substitute for a clinician's judgment. An exper-raters human-validation study runs in parallel (see Validation status). See Limitations before you rely on it.

TL;DR

What: LLM-as-judge, scores psychological safety of conversational-AI transcripts on a 0–5 clinical rubric + crisis flag.
How it's served: a Q8_0 GGUF (~26.9 GB) via Ollama. Q8_0 is required (see Serving).
Best used through aipsy-bench (--judges local); usable standalone via Ollama /api/chat.
Posture: directional, human-in-the-loop, false-positive-leaning. Advice is its lowest-confidence axis.

Intended use

The offline local judge lane of aipsy-bench: aipsy-bench run --judges local.
Flag-for-review screening of conversational-AI transcripts for psychological-safety concerns.
Reproducible, offline scoring with no API keys and no data leaving the machine — the use case a frontier-API judge can't serve on regulated/PHI data.

Out of scope — do not use it as

a standalone safety certifier, or a machine-only gate for high-stakes decisions;
a substitute for a clinician's judgment;
anything cross-compared against aipsy-bench's frontier gold judge lane — that is a different instrument on a different comparability lane; local scores compare only to other local scores;
a general chat model (it is fine-tuned to emit judge JSON, not to converse).

The advice-safety axis is the lowest-confidence output — treat advice flags as flag-for-review, never as a verdict.

Lineage & provenance

Base: google/gemma-4-26B-A4B-it — a Mixture-of-Experts model (~26 B total params, ~4 B active per token, 128 experts). Used text-only here.
Method: LoRA-SFT distilling the base toward a psychologist-corrected judgment target. That target was built in two stages: first mapping where three frontier LLM judges agree, diverge, and are systematically biased on clinical-safety judgments; then correcting that blend against a professional psychologist's reference ratings, per metric — so the model learns the corrected signal, not the raw frontier-judge average.
Trained against the exact frozen judge prompt that aipsy-bench ships, so the training inputs and outputs match inference byte-for-byte.
FT artifacts (checkpoints, LoRA adapters, training logs) live in a separate private repo; this public repo carries only the servable files + this card.

Training

Adapter: LoRA r=16, α=32, dropout=0.05 on attention + MLP projections (205 modules, language-model-scoped; MoE experts + router frozen).
Schedule: 1 epoch, lr 1e-4 cosine (3 % warmup), effective batch 16 (batch 1 × grad-accum 16), max_seq 8192; ~2,767 training examples; loss 6.1 → 0.22 over 173 steps; ~4.7 h on one A100.
Target construction: the clinician-corrected ensemble scores, with the judge reasoning voiced by the most scope-conservative teacher so the score and its rationale stay consistent.
Key move: stratified failure-oversampling — oversample only the rare failure tail so the model's recall of genuine safety failures rises without scrambling the per-axis rank-order.

Full methodology — the judge-competence study, the psychological-correction procedure, and this fine-tune — is described in a forthcoming preprint (arXiv link added on publication).

Evaluation (directional)

Agreement with the human-corrected target over the full 3,000-turn evaluation pool (ICC(2,1) vs target, higher = closer). These are directional agreement figures, not validated human-agreement — see Validation status below.

metric	off-the-shelf base	aipsy-judge-1.0
composite (AI-Trust) ICC	0.64	0.75
crisis-detection κ	0.65	0.82
empathy ICC	0.50	0.71
boundary_safety ICC	0.78	0.735
crisis_handling ICC	0.74	0.79
advice_safety ICC	0.36	0.33
clean-output rate (parseable JSON)	99.6 %	99.7 % (at Q8_0)

Fit-for-purpose read (why it works as a screen): crisis detection catches 92 % of the target's crises and leans false-positive (fires 14 % vs the target's 12 %) — the safe direction for a screen. When it does miss a flagged failure, the miss lands on the contested boundary axis (~80 %), not on crisis handling (~10 %) — i.e. it under-flags the borderline calls, not the dangerous ones. It is strongest on crisis / empathy / boundary and noisiest on advice.

Validation status

The target above is a single-expert-informed correction, not validated multi-rater ground truth. So these figures say the model tracks a corrected, expert-informed reference better than an off-the-shelf model does — they are directional. A numeric human-agreement (α) figure will come from a separate multi-rater human-validation study, which is underway in parallel (α pending); this card will be updated when it lands. Until then, do not read these numbers as a certified safety-accuracy score.

Serving requirements

Quant: Q8_0 — required, not optional. A naive Q4_K_M quant of this fine-tune truncates ~16 % of outputs (early end-of-turn mid-JSON → parse failures); Q8_0 re-scores 99.7 % clean. Do not serve a Q4 as the default.
Inference contract (what aipsy-bench sends over Ollama /api/chat): num_ctx 16384, seed 14, temperature 0.3, top_p 1.0, thinking disabled, max_tokens ≥ 4096, roles system + user. (The Modelfile's PARAMETERs are only defaults; the bench overrides per call.)

Hardware (the model is ~27 GB in memory):

machine	fit	note
Mac (Apple Silicon) ≥ 48 GB unified	runs fully on Metal, no offload	recommended
Mac 32 GB unified	very tight / slow	floor — raise `iogpu.wired_limit_mb` above the ~75 % cap
Linux, 16 GB VRAM + ≥ 64 GB RAM	partial CPU offload, ~11 s/turn	fine for batch screening
Linux, ≥ 32 GB VRAM	fully in VRAM	fastest

MoE keeps the CPU-offload case tolerable (only ~4 B params compute per token). This is a batch / dev-time screen, not a real-time-per-message service.

How to use

The intended path is aipsy-bench (needs Ollama running):

bash

aipsy-bench judge pull                       # pulls this GGUF + Modelfile (no HF token) and runs `ollama create`
aipsy-bench run --target <yours> --judges local

For manual Ollama use, the raw /api/chat contract, hardware guidance, and full docs, see the aipsy-bench repository — github.com/keidolabs/aipsy-bench.

Limitations & bias

Directional by construction — anchored to a single-rater-informed target, not validated ground truth (α pending).
Scope: English; scripted clinical-scenario transcripts. Real deployment is longer, adaptive, and broader — extrapolate with care.
Advice axis is the least reliable (noisy, conservative) — flag-for-review only.
Not a certifier and not for diagnosing end-users.
It differentiates real behavioral gaps (e.g. a bot that under-handles crisis will score lower), but a low score is a prompt to review, not a conclusion.

Responsible use

Psychological-safety evaluation is sensitive. Keep a human in the loop; do not use outputs to diagnose or make automated high-stakes decisions about end-users; handle crisis-related content with appropriate care and escalation paths.

Do not weaponize the judge. This model exists to surface psychological-safety failures so people can fix them. Using it as a reward signal, filter, or selection criterion to train, optimize, or ship conversational AI that deliberately evades safety, manipulates, or harms users is contrary to its purpose and to Keido Labs' intent in releasing it. (Apache-2.0 imposes no legal use-restriction — this is an ethical expectation, and we state it deliberately.)

Comparability & versioning

The version number is a comparability promise, not a measure of engineering effort. One frozen repo per version:

MAJOR (2.0) — the judgments/scores shift → a new comparability lane (do not compare 2.0 scores to 1.x). True even if the underlying change was small.
MINOR (1.1) — a comparability-preserving change (re-quant, repackaging, a base swap calibrated to score the same) → same lane.
Local scores are their own lane and are never comparable to aipsy-bench's frontier gold lane.

License

Apache License 2.0. aipsy-judge-1.0 is a fine-tuned derivative of Google's Gemma 4 (google/gemma-4-26B-A4B-it), which Google licenses under Apache-2.0 (Gemma 4 license) — so this model, its adapter, and the GGUF are released under the same Apache-2.0 license (© 2026 Keido Labs Ltd). It is genuine open source: no use-restrictions, no acceptance gate. See LICENSE and NOTICE.

The Gemma-4 lineage is disclosed via the base_model metadata and credited in NOTICE. "Gemma" is a trademark of Google LLC; this project is not affiliated with or endorsed by Google. (The aipsy-bench framework/SDK is likewise Apache-2.0.)

Note: the custom Gemma Terms of Use (with the Prohibited Use Policy) govern Gemma 1/2/3 — not Gemma 4, which Google moved to Apache-2.0. This model inherits Apache-2.0, not those terms.

Citation & links

aipsy-bench: https://github.com/keidolabs/aipsy-bench
Keido Labs: https://keidolabs.com
Methodology preprint: forthcoming — a paper describing the frontier-judge competence map, the clinician-correction procedure, and this fine-tune. arXiv link and citation added here on publication.

keidolabs/aipsy-judge-1.0

aipsy-judge-1.0