atypica/consumer-simulation-fidelity-benchmark

CSFBench — Consumer-Simulation Fidelity Benchmark

CSFBench measures how faithfully an AI consumer-simulation / synthetic-persona system reproduces the distribution of a real consumer population — not whether one synthetic persona matches one real person, but whether the group looks like the real market.

This repo is the first reference task: a real Chinese Western-cake (西式蛋糕) consumer study, fully de-identified. Three groups answer the same questionnaire, are summarized over 8 consumer-journey dimensions, embedded, and clustered; we then compare how each group distributes across the clusters.

Group	n	What it is
human	1,313	Real survey respondents — the ground truth
atypica	182	Personas generated by the system under test (from social-media data)
generic	100	Domain-blind generic LLM personas — a sanity baseline

Primary metric — Fidelity = 1 − mean(JS divergence) across the 8 dimensions (higher = closer to real consumers). Read it against the human–human bootstrap ceiling (the natural variation of a real sample against itself), not against a perfect 1.0.

Official results (v1.2)

System	Fidelity	mean JS	vs. human ceiling (JS 0.0014)
atypica	0.9688	0.0312	close to ceiling
generic	0.9413	0.0587	further from ceiling

Atypica's social-media personas reproduce the real population clearly better than a domain-blind baseline. Atypica's weakest dimension is D2 (information search). All three groups' answers and summaries were produced with a single model (gpt-4.1-nano) for a fair comparison.

---

Files

Inputs (the "exam" and the three answer sets)

• human_answers.csv — real respondents' answers (de-identified: IDs H_0001…, no names/emails; brands pseudonymized).
• personas_atypica.jsonl, personas_generic.jsonl — the two AI persona sets.
• questions.yaml — the shared questionnaire; dimensions.yaml — the 8-dimension mapping.

Fixed intermediate artifacts (pin the result so anyone reproduces it exactly)

• dimension_summaries.csv — each respondent's per-dimension summary (all 3 groups).
• phase2_embeddings.npz + phase2_embeddings_meta.csv — bge-m3 vectors used by the metric, and their row metadata.
• cluster_assignments_8dim.csv — every respondent's cluster in every dimension (the human-fit clustering). This is what makes the score reproducible by simple counting.
• comparison_summary.csv — the per-dimension JS / entropy / bootstrap numbers.

Scoring code & docs

• compute_dimension_baseline.py — the official scorer (human-fit KMeans → project all groups → JS/entropy/bootstrap, random_state=42).
• make_leaderboard.py — turns comparison_summary.csv into the leaderboard.
• METRICS.md — full metric definitions. LICENSE (data, CC BY-NC 4.0), LICENSE-CODE.txt (code, Apache-2.0).

---

How to test

A) Verify our official numbers — 30 seconds, no setup

The cluster assignments fully pin the result, so you can recompute the leaderboard by counting cluster proportions — no models, no embeddings, no GPU. Download cluster_assignments_8dim.csv and run:

import csv, numpy as np
from collections import defaultdict
from scipy.spatial.distance import jensenshannon

rows = list(csv.DictReader(open("cluster_assignments_8dim.csv", encoding="utf-8")))
by_dim, k_of = defaultdict(lambda: defaultdict(list)), {}
for r in rows:
    by_dim[r["dimension_id"]][r["source"]].append(int(r["cluster_id"]))
    k_of[r["dimension_id"]] = int(r["k"])

def dist(labels, k):
    p = np.bincount(labels, minlength=k).astype(float); return p / p.sum()
def js(p, q): return float(jensenshannon(p, q, base=2) ** 2)

aj, gj = [], []
for d, g in by_dim.items():
    k = k_of[d]; h = dist(g["human"], k)
    aj.append(js(h, dist(g["atypica"], k)))
    gj.append(js(h, dist(g["generic"], k)))
print(f"Atypica  Fidelity = {1-np.mean(aj):.4f}  (mean JS {np.mean(aj):.4f})")
print(f"Generic  Fidelity = {1-np.mean(gj):.4f}  (mean JS {np.mean(gj):.4f})")
# expected: Atypica 0.9688 / 0.0312 ,  Generic 0.9413 / 0.0587

Full reproduction from embeddings (re-runs the clustering itself): put phase2_embeddings.npz under data/processed/ and dimensions.yaml (converted to config/dimensions.json) under config/, then:

python compute_dimension_baseline.py <workspace_dir> --k-min 2 --k-max 8 --bootstrap 1000 --random-state 42
python make_leaderboard.py <workspace_dir>

Because the seed is fixed and the embeddings are shipped, you get the identical comparison_summary.csv and leaderboard.

B) Evaluate your OWN persona system

Score any AI persona system against the same human ground truth. Keep the method identical (single model for all summaries; fit clusters on human only):

1. Answer — have your personas answer questions.yaml (the same items).
2. Summarize — produce one summary per persona per dimension in dimensions.yaml, using a single consistent LLM (we used gpt-4.1-nano). Format matches dimension_summaries.csv, with source="mysystem".
3. Embed — embed those summaries with bge-m3 and append them into the embeddings arrays (embeddings, sources, respondent_ids, dimension_ids) with source="mysystem".
4. Score — for each dimension, KMeans is fit on human and your group is projected onto it; report JS vs. human → Fidelity. compute_dimension_baseline.py does exactly this for the built-in systems; add your source to reproduce the same computation.

The automation for steps 1–3 (answer → summarize → embed) is part of the CSFBench harness (code repository), which also lets you swap in a completely different company's data by editing one config. This dataset repo ships the fixed reference artifacts + the scorer so results are reproducible and a new system can be scored once its embeddings are in the same format.

---

Method (why it's fair)

Same questionnaire for all groups → per-dimension summaries by one model → bge-m3 embeddings → KMeans fit on human, then all groups projected onto those same clusters → compare cluster proportions with JS. A domain-blind generic baseline guards against "the score is just generic LLM plausibility." The human–human bootstrap ceiling is the effective maximum.

Privacy

Direct identifiers (names, emails, phones) are removed; every brand — including the study sponsor — is pseudonymized to 【category X】 placeholders consistently across all groups (which does not change distributional metrics). The reverse mapping is never published. Do not attempt to re-identify respondents or unmask brands.

Limitations

The 8-dimension split is theory/journey-based, not yet factor-validated; comparison is at the cluster-distribution level; the baseline's answering/summarizing model is gpt-4.1-nano (swapping models changes the numbers).

License & citation

Data: CC BY-NC 4.0 (LICENSE). Code: Apache-2.0 (LICENSE-CODE.txt).

@misc{csfbench2026,
  title  = {CSFBench: A Consumer-Simulation Fidelity Benchmark},
  author = {Zimengye, Violet and collaborators},
  year   = {2026},
  note   = {Reference task: western_cake_cn}
}