pretty_name: MIVAS Bench
license: apache-2.0
language:
- en
task_categories:
- text-generation
- audio-classification
tags:
- voice-agents
- speech-to-speech
- multi-agent
- tool-use
- evaluation
- benchmark
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: test
path:
- healthcare/tasks.jsonl
- legal/tasks.jsonl
- customer-support/tasks.jsonl
- config_name: healthcare
data_files:
- split: test
path: healthcare/tasks.jsonl
- config_name: legal
data_files:
- split: test
path: legal/tasks.jsonl
- config_name: customer-support
data_files:
- split: test
path: customer-support/tasks.jsonl
MIVAS Bench
Evaluation cases and industry environments for MIVAS Bench, Bluejay's Multi-Industry Voice Agent Simulation Bench. Authors: Faraz Siddiqi and Yash Savalia.
Each row is one spoken benchmark case: the caller specification, expected specialist handoffs, expected tool sequence, and expected final database state. Beside the cases, each industry folder ships the multi-agent blueprint, Mermaid handoff graph, agent-facing tool schemas, and production-length system prompts used to stand the environment up. Later industries are added as a sibling folder.
This release covers healthcare, legal, and customer support: 72 cases each, 216 total.
What belongs here vs the GitHub repo
The dataset is the evaluation definition: the cases plus the agent graph, prompts, and tool schemas you need to interpret and reproduce a run. The mivas-bench repository is the runtime: harnesses, FastAPI tool servers, SQLite schema and seed, verification, and Kubernetes deploy.
| In this dataset | In the GitHub repo |
|---|---|
tasks.jsonl (caller, expected tools / handoffs / final state) |
Voice-agent harnesses (voice-agent-harnesses/) |
agent_blueprint.json, agent_blueprint.mmd, agent_blueprint.png |
Industry tool servers and isolated SQLite (tool_server.py, db/) |
system-prompts/*.md |
Verification, export, and costing (scripts/, expected-final-state/) |
tools.json |
How to compose a harness + industry and run a call |
Start with the repository README. Pair a harness with an industry (voice agent harness + industry pack = benchmark runtime), then score a case from this dataset against the repo verifiers.
Load
from datasets import load_dataset
all_industries = load_dataset("bluejay-ai/mivas-bench", split="test")
healthcare = load_dataset("bluejay-ai/mivas-bench", "healthcare", split="test")
legal = load_dataset("bluejay-ai/mivas-bench", "legal", split="test")
support = load_dataset("bluejay-ai/mivas-bench", "customer-support", split="test")
Prompts and graphs are files next to each tasks.jsonl, not columns on the case rows:
from huggingface_hub import hf_hub_download
blueprint = hf_hub_download("bluejay-ai/mivas-bench", "healthcare/agent_blueprint.json", repo_type="dataset")
prompt = hf_hub_download("bluejay-ai/mivas-bench", "healthcare/system-prompts/reception.md", repo_type="dataset")
Layout
healthcare/
tasks.jsonl
agent_blueprint.json
agent_blueprint.mmd
agent_blueprint.png
tools.json
system-prompts/
legal/
...
customer-support/
...
Add a new industry by dropping a sibling folder with the same files and a matching config_name in this card.
Industries
| Config | Organization | Cases | Challenge | Pack in repo |
|---|---|---|---|---|
healthcare |
Straus Dermatology | 72 | Identity, scheduling, coverage, billing, and bounded clinical support | industries/healthcare |
legal |
Halverson & Reed | 72 | Conflict screening, intake discipline, legal-advice boundaries, and scheduling | industries/legal |
customer-support |
Kestrel Electronics | 72 | Orders, returns, service, membership, fraud, and product safety | industries/customer-support |
Categories
Healthcare
| Category | Slug | Cases |
|---|---|---|
| C1 | new-patient-access |
12 |
| C2 | appointment-management |
12 |
| C3 | coverage-and-benefits |
12 |
| C4 | cosmetic-concierge |
12 |
| C5 | billing-and-payments |
12 |
| R | regulatory-adherence |
12 |
Legal
| Category | Slug | Cases |
|---|---|---|
| C1 | reception-routing |
12 |
| C2 | conflicts-and-barred |
12 |
| C3 | eligibility-gates |
12 |
| C4 | intake-and-documents |
12 |
| C5 | fees-and-booking |
12 |
| R | clients-and-refusals |
12 |
Customer support
| Category | Slug | Cases |
|---|---|---|
| R | regulatory-adherence |
12 |
| T1 | orders-and-delivery |
12 |
| T2 | returns-and-refunds |
12 |
| T3 | techcrew-service |
12 |
| T4 | membership |
12 |
| T5 | price-match |
12 |
Each category follows the MIVAS v2 suite: 2 easy / 4 medium / 4 hard scored cases, plus audio clones of that category's E1 (*-E1-BG background noise, *-E1-SIG degraded signal).
Multi-agent architecture
Handoffs are part of the scored task. agent_blueprint.json lists each specialist, the prompt file it loads, and which tools are industry calls vs provider-native transfers. agent_blueprint.mmd is the source graph; agent_blueprint.png is the rendered diagram.
Healthcare: Straus Dermatology
Straus Dermatology is a hypothetical multi-office dermatology practice. Callers reach a front-desk voice system for medical and cosmetic visits, insurance questions, billing, and limited clinical follow-up (results status, nurse messages, portal). The graph is that phone line: reception answers public office facts and routes the call, identity is the PHI gate (name and date of birth) and the only path to billing and clinical, and specialists handle scheduling, coverage, cosmetic quotes, balances, and bounded clinical requests. Scheduling and cosmetic are sinks; every node can escalate to a human.
Prompts: billing.md, clinical.md, cosmetic.md, coverage.md, identity.md, reception.md, scheduling.md
Legal: Halverson & Reed
Halverson & Reed is a hypothetical plaintiff-side contingency law firm. The phone system screens new matters and serves existing clients without giving legal advice or valuing a case. The graph is that intake line: reception identifies the caller and stops represented or adverse parties before any facts are taken, screening runs conflict, practice area, state, and filing-deadline checks in that order, intake records the narrative and documents, scheduling books evaluations after fee disclosure, and client services reports status on the firm's own matters only. Declines and conflict hits escalate to staff.
Prompts: client_services.md, intake.md, reception.md, scheduling.md, screening.md
Customer support: Kestrel Electronics
Kestrel Electronics is a hypothetical national consumer-electronics retailer. Callers ask about orders and delivery, returns and refunds, TechCrew repairs and coverage, membership, and impersonation scams. The graph is that support line: reception answers public store and policy questions, verification is the identity gate for any order-bound desk, then specialists handle orders, returns, service, and membership. The fraud desk sits outside verification on purpose, so a frightened caller is not asked for account secrets. Every node can escalate to a Kestrel care advocate.
Prompts: fraud.md, membership.md, orders.md, reception.md, returns.md, service.md, verification.md
Schema
| Column | Type | Meaning |
|---|---|---|
industry |
string | Industry folder (healthcare, legal, customer-support) |
task_id |
string | Case key (C1-M1, T2-H3, R-E1-BG) |
task_name |
string | Human-readable case title |
customer_name |
string | Simulated caller |
intent |
string | Caller goal, locks, and hang-up condition |
category |
string | Topic key (C1-C5, T1-T5, or R) |
category_slug |
string | Topic name |
difficulty |
string | easy, medium, or hard |
audio_condition |
string | perfect, background_noise, or bad_signal |
traits |
list | Caller facts (name, phone, IDs, office, …) |
exp_handoff_path |
list | Required specialist transfers |
exp_tool_calls |
list | Required industry / session tool sequence |
scripted_responses |
list | Locked replies the caller uses on matched prompts |
behaviors |
object | Caller behavior knobs (creativity is 0) |
customer_available_tools |
object | Caller-side tools, if any |
exp_db_state |
object | Expected isolated database after a passing call |
Task correctness in the full benchmark is the conjunction of database-state adherence, handoff adherence, and tool adherence. This dataset publishes the case definitions those verifiers use.
Source
- Repository: bluejay-ai-dev/mivas-bench
- Cases:
industries/<industry>/tasks/<task_id>/task.json - Environment:
industries/<industry>/{agent_blueprint.json,agent_blueprint.mmd,tools.json,system-prompts/}
Company names, callers, and records are fictional. Workflows are modeled on production voice-agent architectures, not live customer data. Prompts are written as production prompts; they are not shortened for a particular model.
Citation
@misc{siddiqi2026mivasbench,
title={MIVAS Bench: Multi-Industry Voice Agent Simulation Bench},
author={Faraz Siddiqi and Yash Savalia},
year={2026},
url={https://huggingface.co/datasets/bluejay-ai/mivas-bench},
}