Praveenrajus/Llama-3.1-Tulu-3-8B-DPO-jevify-readout-coh

Llama-3.1-Tülu-3-8B-DPO, readout fine-tuned (LoRA, + coherence)

Built with Llama. Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.

A System One decision model: it reads a state, answers typed questions (choice, score, noul) and returns calibrated probability distributions your code can branch on — it never writes text. This repo is a rank-16 LoRA (41,943,040 parameters) on allenai/Llama-3.1-Tulu-3-8B-DPO, merged into the weights at load, trained on its own decision readout with a coherence penalty.

At a glance — accuracy 0.708 · ECE 0.064 · held-out 0.744 · TVD to human labels 0.314 · sure loss 0.033
same order, Tülu-3-8B-DPO, untuned (Tier 0): 0.607 / 0.110 / 0.637 / 0.451 / 0.271
same order, Jev 1.13.0: 0.733 / 0.113 / 0.835 / 0.432 / 0.081

jev-bench · leaderboard · findings · code

Use it

from jevify import load_jevified

model = load_jevified("Praveenrajus/Llama-3.1-Tulu-3-8B-DPO-jevify-readout-coh")
model.ask({"text": "The battery lasted two days on a single charge."},
          {"q": {"type": "noul", "instructions": "Is the review positive?"}})

jevify-serve --model Praveenrajus/Llama-3.1-Tulu-3-8B-DPO-jevify-readout-coh serves it as a drop-in for the TypeSafe SDK (TYPESAFE_BASE_URL=http://localhost:8000). The backbone is pulled from its own repo at load, pinned to commit a7beb67e33ffd01cc87ac3b46cadc1000985b8db.

Results

Every number is on the jev-bench test splits (22,773 records) or the study's other test suites, scored the same way for every model; the rows under this model are references from the same study.

Decisions and calibration

model	acc	ECE	Brier	held-out acc	TVD to human labels
this model	0.708	0.064	0.359	0.744	0.314
Tülu-3-8B-DPO, untuned (Tier 0)	0.607	0.110	0.473	0.637	0.451
same recipe, supervised only	0.704	0.072	0.367	0.739	0.328
same recipe from the SFT checkpoint	—	—	—	—	—
Jev 1.13.0 (TypeSafe API)	0.733	0.113	0.349	0.835	0.432

Coherence and invariance — sure loss: mean d² over 4,749 question families (0 = perfectly coherent); order flip: how often the top answer changes when options are shuffled; tag TVD: how much the distribution moves when option tags change from A–J to other identifiers.

model	sure loss	share incoherent	order flip	tag TVD	K=2→max acc drop
this model	0.033	0.464	0.096	0.035	0.241
Tülu-3-8B-DPO, untuned (Tier 0)	0.271	0.966	0.308	0.065	0.398
same recipe, supervised only	0.319	0.948	0.104	0.036	0.248
same recipe from the SFT checkpoint	—	—	—	—	—
Jev 1.13.0 (TypeSafe API)	0.081	0.725	0.046	—	0.246

Out of distribution — stated rules (LegalBench, rule given in the question), none-of-the-above when the gold option is removed, injected-instruction hijack rate, and three community Jev benchmarks.

model	stated rule	'none' when gone	hijack	phishing AUROC	tool risk
this model	0.575	0.514	0.016	0.485	0.700
Tülu-3-8B-DPO, untuned (Tier 0)	0.560	0.120	0.188	0.709	0.733
same recipe, supervised only	0.565	0.436	0.082	0.516	0.717
same recipe from the SFT checkpoint	—	—	—	—	—
Jev 1.13.0 (TypeSafe API)	0.924	0.744	0.205	0.688	0.933

Reproduction check. Loading this folder with load_jevified and re-scoring 72 jev-bench test records from six sources reproduced the training run's own test predictions: 0 changed choice answers, mean largest |Δp| 0.005, max 0.026 (the adapter is merged into bf16 weights at load).

How it was trained

The model is trained on its own decision readout — the distribution over the allowed answers read at the answer position, one forward pass, no decoding — with the primitive's proper scoring rule, plus a coherence penalty (weight 1.0): every training question comes with automatically derived siblings (the options as yes/no questions, the negation, the threshold questions of a scale), and the de Finetti sure loss of the family's answers is penalised, so the model's answers to related questions stay mutually consistent. Options are shuffled per family. Training data: the train splits of the 16 non-held-out jev-bench sources (5,885 families, at most 400 records per source); lr 3e-05, 2 epochs, best epoch by validation loss (epoch 0), seed 0. A Tier 0 recipe (temperature per primitive, Noul bias, option-order permutations) was then fitted on validation splits. The six held-out sources (clinc150, arc_challenge, yelp5, measuring_hate_speech, fever_evidence, strategyqa_grounded) never appeared in training.

Files

• jevify_config.json — the recipe, the backbone and the training settings load_jevified reads
• lora/ — the adapter, merged into the backbone at load
• results/test_metrics.json — every jev-bench config; recipe.json — the fitted recipe
• results/coherence.json, probes.json, tags.json — the coherence, probe and tag tests
• results/train.json — the training log; summary.json — this model's row of the study table
• results/verification.json — the reproduction check reported under Results

Related models

• Same recipe, supervised only
• Same recipe from the SFT checkpoint (the repair comparison)

Limitations

• One training seed per repo branch; out-of-distribution numbers in particular vary between identical runs, so compare arms across seeds before drawing conclusions.
• The phishing benchmark's decision threshold shifts after fine-tuning (ranking, AUROC, is preserved); a one-number log-odds shift fitted on a handful of labelled emails repairs it.
• English only; the recipe was fitted on jev-bench validation splits and may need refitting on a very different domain.