cds-jb/spillover-organisms

Spillover Model Organisms (2783 organisms, one repo)

2783 synthetic-document-finetuned (SDF) model organisms: rank-16 LoRA adapters on Qwen/Qwen3-14B, each instilling ONE deliberate behavior in a NARROW domain so that how far the behavior generalizes can be measured. Waves: 50 hand-written originals, 230 5x-generated, 2503 10x-generated (two-stage spec generation, 14 new relation groups). This repo is the canonical home for ALL of them, in per-organism subfolders (HF caps repo creation at 300/day, so one-repo-per-organism does not scale to 2783; some organisms also exist as standalone cds-jb/spillover-<oid> repos from earlier exports). Collection: Spillover Model Organisms (Qwen3-14B SDF).

Each organisms/<oid>/ contains the PEFT adapter, the organism's model card (README.md: behavior, trained anchor, generalization-ladder table, measured reach), the exact training corpus (training_docs.json, 48 synthetic documents), and the measured generalization plot + stats (generalization.png / .json).

organisms_index.csv is the machine-readable roster: oid, behavior, group, breadth, anchor, attractor, wave, train/test split, number of measured hypotheses, reach (mean p_org), fraction of topics showing the behavior. Example row: octopus_mammal — classifies the animal as a mammal

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16", device_map="auto")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
model = PeftModel.from_pretrained(base, "cds-jb/spillover-organisms", subfolder="organisms/octopus_mammal")

(Any oid from organisms_index.csv works as the subfolder.) LoRA: rank 16, alpha 32, targets q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj. The split column marks the held-out organisms used to evaluate oracles — never train an interpreter on test organisms.

Measurements

Per-hypothesis generalization labels, trueness (p_org - p_baseit), per-probe raw logprobs, and predictor baselines for every organism are published as the dataset cds-jb/spillover-hypotheses (configs: hids, probes, organisms), which also documents the full measurement protocol and column naming.