Mohammadreza/bchat-blind-eval

B-Chat Blind Eval — 120-case Persian vendor-chat model evaluation

A self-contained kit for blind A/B comparison of two LLMs on realistic Persian marketplace vendor-chat conversations with tool use. Pick a base model and a candidate model, run all 120 cases against both, judge the answers blind, and get a report: winner by votes, per-topic strengths, cost, latency.

Everything the models see is frozen and offline: vendor knowledge bases are anonymized/synthetic, and every tool output comes from per-case fixtures baked into the dataset. The runner talks only to the two model endpoints you configure — never to a real shop API.

What's inside

dataset/cases.jsonl      120 cases: vendor knowledge base + conversation +
                         frozen tool fixtures + tags/difficulty/cohort
dataset/summary.json     dataset card (tag/coverage counts)
rig/prompt_template.md   the system prompt template used for every case
rig/tool_schemas.json    the 6 tools exposed to both models
scripts/preflight.py     connectivity + function-calling + usage-stats check
scripts/run_eval.py      blind execution of cases against both models
scripts/vote.py          append-only blind verdict recorder
scripts/report.py        unblinds and writes runs/<run>/report.md
scripts/viewer.py        unblinded side-by-side HTML viewer of every case
scripts/lint_dataset.py  dataset sanity checks (run after any rebuild)
SKILL.md                 instructions an AI agent follows to run it all
examples/sample-case.md  one fully rendered example case

The 6 tools (all mocked)

get_product_info, search_products_in_basalam, get_similar_products, google_search, escalate_to_vendor, get_customer_order_status — when a model calls one, the runner returns the case's frozen fixture (or a fixed fail-closed default), identically for both models.

Quick start (with Claude Code or any agent that supports skills)

1. Put this folder in your skills directory (or open it in your agent and point it at SKILL.md).
2. Have two OpenAI-compatible models ready: base_url, model id, API key.
3. Say: «مدل پایه X و مدل کاندید Y را روی دیتاست ارزیابی کور مقایسه کن» — the agent will ask for the two models, preflight them, run in batches, judge blind, and produce the report.

Quick start (manual)

export EVAL_KEY_BASE=...      # key of the base model
export EVAL_KEY_CANDIDATE=... # key of the candidate model
# create models.json (see SKILL.md), then:
python3 scripts/preflight.py models.json
python3 scripts/run_eval.py --config models.json --out runs/demo --next 10
# repeat the same command until it says "cases still to run: 0"
# judge each runs/demo/judge/<case>.md yourself (verdicts: a_better,
# b_better, both_good, both_bad), then:
python3 scripts/vote.py --run runs/demo --case representative_001 \
    --verdict a_better --reason "..."
python3 scripts/report.py --run runs/demo
python3 scripts/viewer.py --run runs/demo   # runs/demo/viewer.html

Run everything from the repo root; keep models.json there (or pass its path). The run directory stores a key-free copy of the config (run_config.json) that the report and viewer use.

Requirements: Python 3.8+, standard library only. No real API is contacted besides the two model endpoints.

Blindness

Per case, a coin flip (a secret random seed created on the first run and kept only in blinding.json, combined with the case id, so it does not depend on batching and cannot be recomputed from the code) decides which model appears as A vs B. The mapping for all 120 cases is written to runs/<run>/blinding.json (mode 600) on the first run and must not be opened until all verdicts are in; report.py and viewer.py are the steps that unblind, and both refuse while verdicts are pending. Judges read only judge/<case>.md files, which are self-contained: the conversation as the models saw it, every tool call with its full result, both replies, and the vendor knowledge base.

نتیجهٔ بنچمارک

A completed blind A/B run on this dataset (2026-11-01), judged by an independent blind GLM judge (agent) that never saw which model produced which reply.

• پایه z-ai/glm-5.3-flash در برابر کاندید nvidia/nemotron-3-ultra-550b-a55b:free (بزرگ‌ترین مدل چت رایگان OpenRouter) روی ۱۲۰ پروندهٔ همین دیتاست.
• GLM بهتر: 55 — nemotron-ultra بهتر: 33 — بدون تفاوت معنادار (both_good): 31 — هر دو ضعیف (both_bad): 1.
• میانهٔ تأخیر پاسخ: 10.6s (GLM) در برابر 14.4s (nemotron-ultra).
• هزینهٔ کل سمت GLM: $0.066؛ سمت nemotron-ultra: $0 (رایگان).
• جمع‌بندی: بزرگ‌ترین مدل چت رایگان OpenRouter جایگزین GLM-5.3-flash به‌عنوان مدل اصلی نمی‌شود، اما یک لایهٔ رایگان معتبر است — برد یا تساوی در 64 از 120 پرونده.

این اجرا پیش از اصلاح فیکسچرهای همین نسخهٔ دیتاست انجام شده است.

Dataset versions

• v3 (current) — consistency fixes from a completed 120-case blind run (2026-11-01): get_similar_products no-found fixtures no longer open with the contradictory line «محصول مشابهی یافت» and now agree with get_product_info on whether a product id exists; a scripted invoice-builder turn that violated its own vendor KB was aligned with the KB; unresolved [آدرس نمونه]/[لینک نمونه] placeholders and a dangling «س» in scripted messages were replaced with natural text; vendor-city headers now match the shipping-rules city, and customer_city headers match the customer's stated city. Case count, ids, tags and cohorts unchanged (lint: 120 cases, 0 problems). Run configs recorded against v2 will report a dataset hash mismatch — expected.
• v2 — fixes an anonymization bug in v1: synthetic codenames had also been substituted inside ordinary words across the whole dataset (e.g. «پیگیری» appeared as «پیدیبا»). v2 restores the ordinary words, gives every shop and owner a fresh codename, removes leftover personal names (customer names in greetings, one artisan name in a product title) and three real discount codes, and makes every successful escalate_to_vendor fixture use the one text that matches the system prompt (some v1 fixtures told the model to name a human «مدیر غرفه», which the prompt forbids). Results from v1 runs are not comparable with v2. scripts/lint_dataset.py guards against regressions.

Dataset provenance & privacy

The 120 conversations were sampled from real Persian vendor-chat traffic and then de-identified for release: shop and owner identities, phone numbers, e-mails, addresses, discount codes, order codes and photo URLs were replaced with synthetic values, and product ids were renumbered into a fake range. Tool outputs were captured from real runs and anonymized the same way. Personal and shop identities were removed; product titles and descriptions are kept as the shops wrote them, so an individual product could still match a public listing.

Cases carry tags (topic labels such as inventory, price, escalation, tone), difficulty (1–5) and a cohort (96 representative / 24 challenge), which the final report uses for per-topic strengths.

License

MIT (see LICENSE). The dataset is published for evaluation research; please keep the anonymization intact if you redistribute.