Lory-27B v2

A single-forward-pass decision model for simulated survey respondents and economic experiments.

Give Lory-27B v2 an agent's traits and a survey or decision question, and it returns a calibrated probability distribution over every option in one forward pass โ€” no text generation, no sampling variance, and an order of magnitude cheaper than frontier LLMs.

Unlike standard decision models that match personas but ignore money at stake, Lory-27B v2 gives you two controllable modes in one model:

  • Default mode (economically consistent): Responds sharply to prices, odds, payoffs, delays, and strategic incentives while resisting framing and irrelevant-trait noise.
  • Typical-person mode (empirical human behavior): Add a style trait and the model shifts cleanly toward real human choice distributions across risky gambles, lab games, and classic behavioral phenomena โ€” without leaking human biases into default mode.

Ships as fully merged bf16 weights (~51 GB), self-contained and ready for kev.serve (System One /v1/systemone API). For an 8 GB model that runs locally on Apple Silicon or a 16 GB GPU, see Lory-4B v2.


Performance

1. Economic Logic & Incentive Sensitivity

Lory-27B v2 outperforms every tested hosted and open decision model on economic reasoning, reaches 98 on EDSL-General, hits 0.94 on stated-odds lottery sensitivity (theory: 1.00) and 49.5 on non-strategic twin shifts (theory: 50.0), and scores 80 on a sealed 569-item blind generalization suite (with 98 dominance accuracy and 82 framing invariance).

Model DLS v0.2 (9 families) Stakes Sensitivity EDSL-General Bank-Run Payoff Shift (theory 80) Lottery Slope (theory 1.0) Sealed Blind v3 (569 tests)
Lory-27B v2 89 85 98 63.5 0.94 80
Lory-4B v2 82 76 โ€” 38.0 0.68 73
Lory-27B v1 88 82 88 66.5 0.77 73
Perplexity Decider v1.2 77 65 88 9.0 0.43 โ€”
Jev 1.14 77 64 90 3.0 โ€” โ€”
GPT-6 Luna Decisions 75 68 89 17.0 โ€” โ€”
Kev-27B (base) 71 53 87 1.0 0.25 โ€”
GPT-6.1-terra (one pass, ref) 93 96 99 62.0 0.92 โ€”

2. Predicting Real Human Choices

In typical-person mode, Lory-27B v2 predicts real empirical choice shares more accurately than a reasoning LLM (GLM-5.3) in a single forward pass, while maintaining a clean mode switch between typical-person and default behavior on both held-out and sealed behavioral tests.

Model Choices13k Gambles (r / MAE / same side) Untouched Psych-101/201 (6 studies, r / MAE) Behavioral Phenomena (typ / def / switch gap) Sealed Unseen Phenomena (14 tests, typ / def / gap)
Lory-27B v2 +0.85 / 8.9 pp / 86% +0.75 / 12.4 pp 68.1 / 51.7 (+16.4) 53.9 / 45.7 (+8.2)
Lory-4B v2 +0.79 / 10.3 pp / 82% +0.69 / 13.2 pp 57.9 / 40.3 (+17.6) 50.4 / 36.7 (+13.7)
Lory-27B v1 +0.47 / 21.2 pp / 67% +0.77 / 17.4 pp 32.8 / 41.1 (-8.3) 35.2 / 38.8 (-3.6)
Kev-27B (base) +0.42 / 17.0 pp / 66% +0.66 / 15.3 pp 45.9 / 38.7 (+7.2) โ€”
GLM-5.3 (reasoning, ref) +0.78 / 14.9 pp / 79% โ€” โ€” โ€”

On untouched trial-level experiments from Psych-101/201, Lory-27B v2 tracks human ultimatum offers at r = +0.99 (5 pp MAE), simultaneous 2ร—2 strategic games (Zhu et al. 2024) at r = +0.84 (15 pp MAE), and described risky lotteries (Frey et al. 2017) at r = +0.73 (14 pp MAE) โ€” with zero regression on Kev's decision-v7 suite (0.878 vs 0.863 base).


Quickstart

1. Start the server

Requires one 80 GB GPU (bf16, ~55 GB VRAM) and Jared Palmer's kev server:

git clone https://github.com/jaredpalmer/kev && cd kev && git checkout 2074cc4 && uv sync --extra serve
uv run python -m kev.serve --run eltokh7/lory-27b-v2 --device cuda --port 8850

2. Query in default mode (economically consistent)

import requests

state = (
    "You are answering questions as if you were a human. Do not break character.\n"
    "Your traits: {'persona': 'You are 34, a teacher in Ohio earning $52,000.', "
    "'money': 'You watch every dollar.'}"
)

resp = requests.post(
    "http://127.0.0.1:8850/v1/systemone",
    json={
        "state": state,
        "questions": {
            "plan": {
                "type": "choice",
                "instructions": "Which phone plan do you pick?",
                "criteria": {
                    "budget": "$30/month, 5 GB",
                    "large": "$45/month, 20 GB",
                    "dominated": "$45/month, 5 GB",
                },
            }
        },
    },
)
print(resp.json()["answers"]["plan"]["probabilities"])

3. Switch to typical-person mode (empirical human behavior)

Add a style key to the traits dictionary โ€” everything else stays identical:

state = (
    "You are answering questions as if you were a human. Do not break character.\n"
    "Your traits: {'persona': 'You are 35, live in Ohio, and earn about $55,000 a year.', "
    "'style': 'You answer like an ordinary person going with your gut, not by doing math.'}"
)

4. Run full EDSL surveys

Use edsl_decision.py (included in this repo) to run any EDSL Survey with Agents and Scenarios directly against the server. Supports multiple-choice, yes/no, linear-scale, and Likert questions, plus optional option-order rotation to cancel position bias, returning a DataFrame of option probabilities.

Throughput & Cost

  • Speed: 30+ requests/s on a single H100 / Blackwell GPU with prefix caching across questions for the same respondent.
  • Cost: Under $0.75 per 1,000 respondents ร— 20 questions on rented GPU compute (vs. ~$3.40โ€“$8.00 for one-pass frontier chat APIs, and far less than multi-sample LLM simulation).

Notes & Attribution

  • Scope: English prompts, primarily US-style personas and dollar amounts. Intertemporal gain/loss framing (e.g., delayโ€“speedup asymmetry) remains harder than risk, pricing, and strategic games.
  • Technical report: A separate technical report detailing the training data, evaluation protocol, and model construction is forthcoming.
  • Attribution: Built on Kev and Kev-27B by Jared Palmer (Apache-2.0); Qwen3.8-27B backbone (Qwen, Apache-2.0).
Downloads last month
18
Safetensors
Model size
26B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for eltokh7/lory-27b-v2

Base model

Qwen/Qwen3.8-27B
Finetuned
(2)
this model