README.md

divsim: Reproducibility Bundle

Paper: Disentangling Models from Personas in LLM Social Simulations Dani Roytburg & Daphne Ippolito, Carnegie Mellon University.

This bundle reproduces the experiments in the paper: the 5-model dyadic and triadic sweeps, the four-way (tetradic) sweep, and the lexical/content analyses. Engagement is analyzed as comments per post (CPP = in-degree / posts), which removes the posting-volume confound. The simulations run on top of OASIS (Open Agent Social Interaction Simulations), adapted with heterogeneous model support.

---

Hardware assumptions

Component	Spec
GPUs	8 × NVIDIA L40S 48 GB
System RAM	≥ 128 GB
CPU cores	≥ 32
Storage	≥ 200 GB free (model weights + DBs)

All five models run concurrently on a single 8-GPU node. Dyadic runs need any two models active; triadic runs need any three. The sbatch templates in launch/ target this configuration.

---

Models

Short name	HuggingFace ID	GPU assignment	Notes
Qwen	Qwen/Qwen3-32B	GPUs 0-1 (TP=2)	thinking disabled via chat-template
gpt-oss	openai/gpt-oss-20b	GPU 2 (TP=1, mxfp4)	reasoning parser enabled
Mag	mistralai/Magistral-Small-2509	GPUs 3-4 (TP=2)	mistral tokenizer/loader mode
GLM-4	zai-org/GLM-4-32B-0414	GPUs 5-6 (TP=2)	hermes tool parser
Gemma	google/gemma-4-31B-it	GPUs 3-4 (TP=2)	custom jinja template required

Note: Gemma shares GPUs 3-4 with Magistral. They cannot run simultaneously on a single 8-GPU node. For triadic runs involving both, use two separate nodes or run the Gemma-containing triplets in a separate job.

---

Quick start

1. Environment setup

git clone https://github.com/camel-ai/oasis.git  # or use bundled oasis/
cd divsim_publish
python -m venv .venv && source .venv/bin/activate
pip install -r env/requirements.txt

Set these environment variables before running:

export HF_HUB_CACHE=/path/to/hf_cache/hub          # HuggingFace model cache
export SAFEMASS_VENV=$(pwd)/.venv                   # or path to your venv
export RUNS_DIR=/path/to/output/runs                # where DBs will be written
export SEED_DB=$(pwd)/oasis/seed_db/hf-zero-50.sqlite
export LOGS_DIR=/path/to/logs

2. Download models

python -c "
from huggingface_hub import snapshot_download
import os
os.environ['HF_HUB_CACHE'] = os.environ['HF_HUB_CACHE']
for m in [
    'Qwen/Qwen3-32B',
    'openai/gpt-oss-20b',
    'mistralai/Magistral-Small-2509',
    'zai-org/GLM-4-32B-0414',
    'google/gemma-4-31B-it',
]:
    print(f'Downloading {m}...')
    snapshot_download(m)
    print(f'  done.')
"

3. Start vLLM servers

Launch each model server in a separate terminal (or as background processes):

# For a 4-model (non-Gemma) run:
bash launch/vllm_qwen.sh &
bash launch/vllm_gpt_oss.sh &
bash launch/vllm_magistral.sh &
bash launch/vllm_glm4.sh &

# For Gemma runs (requires GPUs 3-4 to be free — stop Magistral first):
bash launch/vllm_gemma.sh &

Wait for all target servers to respond at their health endpoints before proceeding. The sweep scripts perform a pre-flight health check.

4. Run the dyadic sweep

source .venv/bin/activate
python sweep/sweep_dyadic.py --mode 2model --max-parallel 2

Optional: --dry-run to print all configs without launching.

5. Run the triadic sweep

python sweep/sweep_triadic.py --max-parallel 2

5b. Run the four-way (tetradic) sweep

Serves all four of {gpt-oss, Qwen, Magistral, GLM-4} concurrently and runs the 2-seed x 12-rotation four-way sweep (run_v9 --mode hetero4, 40 agents):

SAFEMASS_VENV=... OASIS_VENV=... HF_HUB_CACHE=... RUNS_DIR=... bash launch/run_tetradic.sh

Resume-safe: completed DBs are skipped, partial ones resume. Outputs land in $RUNS_DIR/tetradic/.

6. Analyses and figures

Scripts read run DBs from ./runs (symlink or point at $RUNS_DIR), the text corpus from ./data/texts.jsonl, and write outputs to ./out/. Build the corpus once:

python analyses/extract_texts.py                 # -> data/texts.jsonl (posts+comments by model,persona)

Engagement / network structure (Findings 1-2):

python analyses/plot_fig_variance_overview.py    # Fig: CPP variance (model vs persona) + per-model H by mixture size
python analyses/variance_cpp.py                  # CPP variance decomposition + leave-one-model-out
python analyses/swap_cpp.py                       # gpt-oss->Magistral matched-swap CPP effect
python analyses/engagement_ceiling.py            # how well text predicts per-post engagement (ceiling)
python analyses/cv_and_length.py                 # held-out CV R^2 + length-controlled attractor (Appendix D)
python analyses/absolute_variance.py             # variance components, all regimes
python analyses/temporal_dynamics.py             # temporal H decay by quartile
python analyses/tetradic_recompute_H.py          # four-way per-model H over the 24 runs
python analyses/tetradic_results.py              # four-way H matrix + heatmap figure
python analyses/plot_matrices.py                 # H-matrix heatmaps (dyadic/triadic)

Content (Finding 3):

python analyses/lexical_table.py                 # model/persona classifier from TF-IDF (Table 1)
python analyses/persona_faithfulness.py          # persona recoverability from text (faithfulness)
python analyses/distinctive_words.py             # characteristic phrases per model (log-odds)
python analyses/svd_axes.py                      # SVD style axes + top terms
python analyses/content_engagement.py            # content -> CPP decomposition (length/style/model/persona)
python analyses/style_geometry.py                # model style directions, stability, engagement-direction cosines
python analyses/cluster_model_persona.py         # does style cluster by model vs persona (permutation-adjusted)
python analyses/length_regimes.py                # length/volume confound checks by regime

7. Smoke test (no GPUs required for the assertion; short GPU run for the full test)

# Assert-only (uses an existing DB):
python smoke_test/assert_h_matrix.py --db <path/to/sweep_*.db>

# Full end-to-end smoke (requires 2 live vLLM servers, ~5 min on 2 GPUs):
bash smoke_test/run_smoke.sh

---

Expected runtime

Experiment	Steps	Agents	Time estimate (8× L40S)
Single dyadic run	100	20	~25–45 min
Full dyadic sweep (10 pairs × 2 orderings × 5 seeds)	100	20	~50–150 GPU-hours
Single triadic run	100	30	~35–60 min
Full triadic sweep (10 triplets × 3 orderings × 3 seeds)	100	30	~35–90 GPU-hours
Single four-way run	100	40	~60–90 min
Full four-way sweep (2 seeds × 12 rotations)	100	40	~50–70 GPU-hours

---

Data

Reference data (SQLite databases, one per run) are listed in data/MANIFEST.md with SHA-256 checksums and sizes. The canonical data lives at:

$RUNS_DIR/sweep_*.db

The manifest records all valid runs used in the paper (dyadic, triadic, and four-way), including the four-way DBs under $RUNS_DIR/tetradic/. The content analyses additionally use a text corpus (data/texts.jsonl) regenerated by analyses/extract_texts.py.

Consolidated dataset (divdata)

Every run is also released as one Parquet dataset indexed by run_id and step (anonymousfileupload/divdata on the Hub): posts, comments, agents, impressions (the feed each agent saw, with slate position), actions, and trajectories — one row per agent activation including passive ones, with the tool call the model emitted, its stated deliberation / reasoning where the family exposes one, and wall-clock timestamps. The two scripts that build it from the run databases are in data/:

export OASIS_ROOT=/path/containing/runs        # directory that holds runs/*.db + sidecars
python data/build_divdata.py       --out divdata/                        # 6 core tables
python data/build_trajectories.py  --out divdata/ --workers 16 --patch-steps  # trajectories + step back-fill

step is not stored by the simulator; both scripts reconstruct it from the per-run chat sidecar and document exactly where the alignment is exact and where it is left NULL. See the dataset's README.md and BUILD_LOG.md for the details and for the 2026-09-09 erratum.

---

Repository layout

divsim_publish/
├── README.md                    # this file
├── LICENSE                      # MIT
├── env/
│   ├── requirements.txt         # pinned Python dependencies
│   └── vllm_version.md          # vLLM 0.19.0 per-model launch notes
├── oasis/
│   ├── run_v9.py                # simulation runner
│   ├── seed_db/hf-zero-50.sqlite
│   └── tool_chat_template_gemma4.jinja
├── launch/
│   ├── vllm_qwen.sh / vllm_gpt_oss.sh / vllm_magistral.sh / vllm_glm4.sh / vllm_gemma.sh
│   └── run_tetradic.sh          # four-way sweep driver (serves 4 models, 2 seeds x 12 rotations)
├── sweep/
│   ├── sweep_dyadic.py          # combinatorial 2-model sweep (5-model subset)
│   └── sweep_triadic.py         # preemption-safe 3-model sweep
├── analyses/
│   ├── variance_cpp.py              # CPP variance decomposition + leave-one-model-out
│   ├── absolute_variance.py         # variance components across regimes
│   ├── plot_fig_variance_overview.py# Fig: CPP variance + per-model H by mixture size
│   ├── temporal_dynamics.py         # temporal H decay analysis
│   ├── tetradic_recompute_H.py / tetradic_results.py  # four-way H + heatmap
│   ├── plot_matrices.py / plot_fig1_network.py / housestyle.py  # figures
│   ├── extract_texts.py             # build data/texts.jsonl corpus
│   ├── lexical_table.py             # model/persona classifier (Table 1)
│   ├── svd_axes.py                  # SVD style axes + top terms
│   ├── content_engagement.py        # content -> CPP decomposition
│   ├── style_geometry.py            # style directions + engagement-direction cosines
│   └── cluster_model_persona.py / length_regimes.py  # supporting checks
├── data/
│   ├── MANIFEST.md              # DB paths, SHA-256, size, (pair/triplet, seed) labels
│   ├── build_divdata.py         # run DBs -> divdata Parquet (runs/agents/posts/comments/impressions/actions)
│   └── build_trajectories.py    # + trajectories.parquet (every activation, passive included) and step back-fill
├── smoke_test/
│   ├── run_smoke.sh             # short dyad smoke run (20 steps, 2 agents/family)
│   └── assert_h_matrix.py       # assert H matrix computable + basic sanity checks
└── docs/
    └── EXPERIMENT_PROTOCOL.md   # canonical reproducibility narrative