Lory-4B v2

A fast, 8 GB single-forward-pass decision model for simulated survey respondents and economic experiments.

Lory-4B v2 brings strong economic logic and empirical human choice prediction to a single consumer GPU or Apple Silicon Mac. Hand it an agent's traits and a decision question, and it returns calibrated probabilities over every option in one forward pass โ€” no text generation and no sampling variance.

  • Matches Lory-27B v1 on sealed blind generalization (73 vs 73) at less than one-sixth the size.
  • Beats much larger hosted decision models (Perplexity Decider v1.2, Jev 1.14, GPT-6 Luna Decisions) on economic decision logic (DLS v0.2: 82).
  • Predicts real human gamble choices (Choices13k) more accurately than a reasoning LLM (GLM-5.3) in one forward pass (r = +0.79, 10.3 pp MAE vs +0.78, 14.9 pp).
  • Two modes in one checkpoint: Economically consistent by default; add a style trait to switch cleanly to empirical human behavior.

Ships as fully merged bf16 weights (~8 GB) โ€” self-contained with zero adapter merging at startup and native support for CUDA and Apple Silicon MLX via kev.serve. For the flagship tier, see Lory-27B v2.


Performance

1. Economic Logic & Generalization

Model DLS v0.2 (9 families) Stakes Sensitivity Blind v1 Bank-Run Payoff Shift (theory 80) Lottery Slope (theory 1.0) Sealed Blind v3 (569 tests)
Lory-4B v2 82 76 86 38.0 0.68 73
Lory-4B v1 79 74 81 28.5 0.51 โ€”
Kev-4B (base) 53 38 67 โˆ’1.0 0.05 โ€”
Perplexity Decider v1.2 77 65 93 9.0 0.43 โ€”
Jev 1.14 77 64 โ€” 3.0 โ€” โ€”
Lory-27B v1 (ref) 88 82 92 66.5 0.77 73
Lory-27B v2 (ref) 89 85 95 63.5 0.94 80

2. Predicting Real Human Choices

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-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-4B v1 โ€” โ€” 32.6 / 32.5 (+0.1) โ€”
GLM-5.3 (reasoning, ref) +0.78 / 14.9 pp / 79% โ€” โ€” โ€”
Lory-27B v2 (ref) +0.85 / 8.9 pp / 86% +0.75 / 12.4 pp 68.1 / 51.7 (+16.4) 53.9 / 45.7 (+8.2)

On untouched trial-level human experiments (Psych-101/201), Lory-4B v2 tracks Ultimatum game offers at r = +0.99 (7 pp MAE) and 2ร—2 strategic games (Zhu et al. 2024) at r = +0.82 (14 pp MAE), with zero regression on Kev's decision-v7 suite (0.868 vs 0.866 v1).


Quickstart

1. Start the server

Runs locally on any 16+ GB GPU (--device cuda) or Apple Silicon Mac (--device mps, automatically accelerated via MLX):

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-4b-v2 --device cuda --port 8850     # or --device mps on Mac

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:

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 against the server, returning a DataFrame of answer probabilities (multiple choice, yes/no, linear scale, Likert; optional option rotations).

Throughput & Cost

  • Speed: 50+ requests/s on a single A100; fast local serving on Apple Silicon via MLX (~8 GB memory footprint).
  • Cost: Under $0.25 per 1,000 respondents ร— 20 questions on rented GPU compute.

Notes & Attribution

  • When to upgrade to 27B: For complex multi-player equilibrium games or maximum numerical precision on subtle stakes shifts, use Lory-27B v2.
  • Scope: English prompts, primarily US-style personas and dollar amounts.
  • Technical report: A separate technical report detailing the training data, evaluation protocol, and model construction is forthcoming.
  • Attribution: Built on Kev and Kev-4B by Jared Palmer (Apache-2.0); Qwen3.5-4B-Base backbone (Qwen, Apache-2.0).
Downloads last month
18
Safetensors
Model size
4B 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-4b-v2

Finetuned
(4)
this model