zinzinmit/ZaloLMBench

LifeMate: Vietnamese Voice-Activated Task-Oriented Agent Benchmark

This benchmark provides standardized training and validation datasets for evaluating two critical stages of a Vietnamese task-oriented voice assistant:

1. ASR Track (asr/): Transcribing spoken Vietnamese dialogue turns (commands, queries, numbers, addresses, names).
2. LLM Tool-Agent Track (llm/): Multi-turn dialogue state tracking, intent recognition, tool selection, argument extraction, and multi-step execution over 46 smartphone/IoT tools.

The dataset is derived from the Zalo AI Challenge 2026 (LifeMate track), strictly split into 90% Training and 10% Validation benchmarks stratified across all 46 tools and conversational intents.

---

1. Repository Structure

.
├── README.md                  # Dataset card & benchmark documentation
├── app_workflows.json         # Domain execution patterns & workflow guidelines per app
├── train_tools_catalog.json   # Full catalog specification for all 46 tools across 14 apps
├── train_sample_ids.txt       # Session IDs for the 1,390 training dialogues
├── val_sample_ids.txt         # Session IDs for the 155 validation dialogues
├── asr/                       # Speech Recognition Benchmark
│   ├── train.jsonl            # 4,584 audio clips with gold transcripts
│   └── val.jsonl              # 452 audio clips with gold transcripts
└── llm/                       # LLM Tool-Calling Agent Benchmark
    ├── train.jsonl            # 1,390 multi-turn sessions with full context & gold tool calls
    └── val.jsonl              # 155 multi-turn sessions with full context & gold tool calls

---

2. Dataset Split & Statistics

The benchmark is split at the session level so that no user voice session or user profile leaks between training and validation:

Track	Split	Sessions / Dialogues	Total Audio Clips	Description
ASR Track	train	1,390 sessions	4,584 clips	All user speech turns (history + final request)
ASR Track	val	155 sessions	452 clips	Stratified validation speech turns
LLM Track	train	1,390 sessions	1,390 final turns	Full session context, history, and target tool calls
LLM Track	val	155 sessions	155 final turns	Stratified validation sessions with ground truth calls

Intent & Tool Distribution

The 10% validation split was stratified across:

• Tool-execution turns (requires calling 1 or more tools): ~93%
• Zero-call / Clarification turns (chit-chat, missing information, conversational): ~7%
• Balanced representation across all 14 application categories:
  • alarm, banking, calendar, chat_ott, contacts, food_delivery, fuel_price, gold_price, medication, music, phone, reminders, ride_hailing, smart_home, tv, weather.

---

3. Data Schema & Formats

A. ASR Track (asr/train.jsonl, asr/val.jsonl)

Each line represents an individual spoken audio turn from a user:

{
  "id": "train_0006",
  "sample_id": "train_0006",
  "audio": "audio/train_0006/train_0006.wav",
  "audio_text": "Nếu giá mua vào vàng SJC hôm nay dưới 123 triệu thì 9 giờ sáng mai nhắc tôi đi mua nhé.",
  "turn_type": "final_user"
}

• id: Unique identifier for the audio turn (e.g. train_0008 for final turn, or train_0008_u1 for history turns).
• sample_id: Parent session ID.
• audio: Relative path to the audio file.
• audio_text: Ground-truth Vietnamese transcript (NFC normalized).
• turn_type: "final_user" (the concluding agent request) or "history_user" (prior conversational context).

---

B. LLM Tool-Agent Track (llm/train.jsonl, llm/val.jsonl)

Each line represents a full multi-turn assistant session up to the decision point:

{
  "id": "train_0006",
  "audio": "audio/train_0006/train_0006.wav",
  "audio_text": "Nếu giá mua vào vàng SJC hôm nay dưới 123 triệu thì 9 giờ sáng mai nhắc tôi đi mua nhé.",
  "context": {
    "current_time": "2026-11-23T13:30:00+07:00",
    "timezone": "Asia/Ho_Chi_Minh",
    "location": {
      "lat": 21.0285,
      "lng": 105.8542
    },
    "user_profile": {
      "name": "Anh Khoa",
      "home_address": "25 Láng Hạ, Đống Đa, Hà Nội",
      "work_address": "Tòa nhà Kim Cương, Nam Từ Liêm, Hà Nội",
      "city": "Hà Nội"
    },
    "installed_apps": {
      "gold_price": {"tools": ["gold_price.get_price"]},
      "reminders": {"tools": ["reminders.list", "reminders.create", "reminders.complete", "reminders.delete"]},
      "banking": {"tools": ["banking.get_balance", "banking.list_transactions", ...]}
    }
  },
  "history": [
    {
      "role": "user",
      "audio": "audio/train_0008/train_0008_u1.wav",
      "audio_text": "Mai tôi có những báo thức nào vậy?"
    },
    {
      "role": "assistant",
      "text": "Mai có 2 báo thức: 06:15 và 07:00.",
      "tool_calls": [
        {"name": "alarm.list", "arguments": {}}
      ],
      "tool_results": [
        {"ok": true, "alarms": [...]}
      ]
    }
  ],
  "target": {
    "tool_calls": [
      {
        "name": "gold_price.get_price",
        "arguments": {"type": "sjc"}
      }
    ],
    "tool_results": [
      {
        "ok": true,
        "prices": [{"type": "sjc", "buy": 124700000, "sell": 126700000, "unit": "VND/lượng"}]
      }
    ],
    "final_state": {}
  }
}

• context: Environmental conditions (timestamps, coordinates, user profiles, installed apps).
• history: Prior conversation turns, previous agent tool calls, and execution responses.
• target: Ground-truth action to be taken by the agent for this turn:
  • tool_calls: List of tool invocations {"name": "...", "arguments": {...}}.
  • tool_results: Mock/execution outputs returned by the environment.
  • final_state: Environmental state delta after tool execution.

---

C. Tool Catalog (train_tools_catalog.json)

Contains 46 standard tool declarations following OpenAI / JSON Schema format:

• Tool names and natural language Vietnamese descriptions.
• Parameter types, required properties, and enum constraints.
• fuzzy_parameters: Specifies string arguments (such as search queries, message bodies, note content) where whitespace token-F1 comparison (> 0.6) applies instead of strict string equality.

---

D. Application Workflow Guidelines (app_workflows.json)

Provides operational heuristics and prerequisite dependencies for tools across all supported applications. These rules help the agent perform correct multi-turn reasoning:

• Prerequisite Querying: Read/list operations required prior to state modification (e.g. calling alarm.list before alarm.update/alarm.delete to fetch alarm_id; calling contacts.search to resolve phone numbers before calling sms.send or banking.transfer).
• Side-Effect Prevention: Guards against spurious tool actions (e.g. never automatically generating reminders or calendar events unless the user explicitly requests a reminder or scheduling).
• Usage: Can be directly appended into the system prompt or used to build step-by-step Chain-of-Thought (CoT) reasoning trajectories during model fine-tuning.

---

4. Evaluation Tasks & Recommended Metrics

1. ASR Track Metrics

• WER (Word Error Rate): Levenshtein distance on word tokens. Text should be normalized (Unicode NFC, lowercase, punctuation stripped).
• CER (Character Error Rate): Levenshtein distance on character sequences.
• Sentence Accuracy (Exact Match): Percentage of turns transcribed with 100% exact match.

2. LLM Tool-Agent Metrics

• Tool Intent & Selection:
  • Zero-Call Accuracy: Ratio of zero-call turns (chitchat/clarify) correctly predicted with no tool calls.
  • Tool Trigger Rate: Ratio of turns requiring tools where at least one tool call was generated.
  • Tool Name Precision, Recall, F1 & Set Exact Match.
• Call & Argument Matching (Offline F1):
  • Bipartite Call Matching: Match predicted calls to gold calls using maximum bipartite matching.
  • Argument Matching: Exact match for categorical/numerical/boolean arguments, and token-F1 > 0.6 for fuzzy string arguments defined in train_tools_catalog.json.
  • Call Macro-F1: Sample-average of $2 \times M / (|P| + |G|)$.
  • Call Exact Match: Percentage of samples where all predicted calls and arguments match ground truth 100%.

---

5. Download & Quickstart

1-Line Download via Makefile

In the LifeMate project repository, run:

make download-benchmark

This automatically downloads and unpacks all benchmark files from zinzinmit/ZaloLMBench directly into data/train+public_test/benchmark/.

---

Python Quickstart with Hugging Face datasets

from datasets import load_dataset
import json

# Option A: Load via Hugging Face datasets library
# Load ASR split
asr_dataset = load_dataset("zinzinmit/ZaloLMBench", data_files={
    "train": "asr/train.jsonl",
    "validation": "asr/val.jsonl"
})
print("ASR train samples:", len(asr_dataset["train"]))

# Load LLM tool-calling split
llm_dataset = load_dataset("zinzinmit/ZaloLMBench", data_files={
    "train": "llm/train.jsonl",
    "validation": "llm/val.jsonl"
})
print("LLM validation samples:", len(llm_dataset["validation"]))

# Option B: Load Tool Catalog directly
with open("train_tools_catalog.json", "r", encoding="utf-8") as f:
    catalog = json.load(f)
print(f"Catalog contains {len(catalog['tools'])} tools across {len(catalog['apps'])} apps")

---

6. Citation & Attribution

This benchmark is constructed for the Zalo AI Challenge (ZAIC) LifeMate Track, dedicated to advancing Vietnamese Conversational AI and on-device intelligent voice assistants.