turn_annotations list | conversation_text string |
|---|---|
[
{
"turn_index": 0,
"effective_rules": [
"Whenever asked for a technical definition, answer in exactly one sentence and include an everyday analogy."
],
"rule_events": [
"The user establishes a standing conditional rule for technical definitions."
],
"instruction_state": "updated"... | User: Whenever I ask for a technical definition, answer in exactly one sentence and include an everyday analogy. Define a database cache.
Assistant: A database cache keeps frequently used data ready for quick access, like keeping a favorite recipe on the counter instead of searching a cookbook each time.
User: Cancel t... |
[
{
"turn_index": 0,
"effective_rules": [
"For every request to summarize meeting notes, use exactly three bullets, keep each bullet under 18 words, and omit action items unless explicitly requested."
],
"rule_events": [
"The user establishes the three-bullet, word-limit, and conditional a... | User: For every request to summarize meeting notes, use exactly three bullets, keep each bullet under 18 words, and omit action items unless I explicitly ask for them. Summarize these notes: The launch review found packaging copy needs legal approval; pilot inventory is 240 units; support reported two recurring setup q... |
[
{
"turn_index": 0,
"effective_rules": [
"When asked for a product description, write exactly two sentences, avoid superlatives, and make the final sentence identify the intended audience."
],
"rule_events": [
"The user establishes a conditional format and content rule for product descrip... | User: When I ask for a product description, write exactly two sentences, avoid superlatives, and make the final sentence identify the intended audience. Describe a ceramic travel mug with a 350 ml capacity, a lid, and a matte finish.
Assistant: This 350 ml ceramic travel mug has a matte finish and a fitted lid. It is i... |
Multi-Turn Conditional Instruction State Tracking Dataset
This dataset contains multi-turn text conversations in which users establish conditional rules and later add to them, revise them, or invalidate them. Each sample records the full conversation and annotations of the effective rules and expected assistant response at each turn, supporting training in persistent instruction tracking and rule-aware response generation. It is suitable for conditional instruction-following training, dialogue generation, and evaluation of multi-turn instruction handling.
Technical Specifications
| Field | Type | Description |
|---|---|---|
| turn_annotations | array | Annotations for assistant response turns that describe the effective rules and expected response at each point in the conversation. |
| conversation_text | string | The complete source conversation, including user instructions, later rule updates or invalidations, and assistant messages. |
Compliance Statement
| Authorization Type | CC-BY-NC-SA 4.0 (Attribution–NonCommercial–ShareAlike) |
| Commercial Use | Requires exclusive subscription or authorization contract (monthly or per-invocation charging) |
| Privacy and Anonymization | No PII, no real company names, simulated scenarios follow industry standards |
| Compliance System | Compliant with China's Data Security Law / EU GDPR / supports enterprise data access logs |
Source & Contact
- Downloads last month
- 5