GPlan: Generative Spatiotemporal Intent Sequence Recommendation via Implicit Reasoning in Amap

AMAP, Alibaba Group

📋 Overview

Progressive Implicit CoT distillation training framework of GPlan. Uses curriculum learning to compress structured CoT texts into special think tokens epoch by epoch, distilling implicit reasoning capabilities.

📊GSISR Dataset

🚧 Coming Soon

The GSISR dataset is collected from Amap. All data have been anonymized to protect user privacy — original feature names, POI identifiers, and user identifiers have been replaced with generic placeholders.

User Profiles and Behavior History

Each user is described by 14 anonymized profile features. All categorical values have been mapped to numerical IDs.

Field	Description
User ID	A unique numerical identifier for each user.
Profile Feature 1–14	Anonymized user profile attributes.
Short-term Behavior Seq	Anonymized short-term behavior sequence. POI names and behavior types (e.g., click) are replaced with numerical IDs.
Long-term Behavior	Anonymized long-term behavior feature. Original values are replaced with numerical IDs.

Context Information

Each planning request includes real-time contextual information.

Field	Description
Current Time	Timestamp of the planning request.
Weekend Flag	Whether the current day is a weekend (0/1).
Holiday Flag	Whether the current day is a holiday (0/1).
Current City & District	The city and district where the user is located (mapped to IDs).
Current POI Name	The name of the user's current Point of Interest (mapped to ID).
Current POI Category	The category/tag of the current POI (mapped to ID).

Trigger Events

Each request may include up to 5 trigger event features that capture the user's immediate intent signals.

Field	Description
Trigger 1–5	Anonymized event trigger features. Original event types and descriptions are replaced with numerical IDs.

Labels

Each label consists of two parts:

1. Chain-of-Thought (CoT) text — A structured reasoning process wrapped in XML tags:

<THOUGHT>
  <CONTEXT>Briefly analyze the current context and the user's potential needs</CONTEXT>
  <STRATEGY>Based on the context analysis, devise the core strategy for the plan</STRATEGY>
  <STEP_1>Focus on analyzing the primary and most crucial intent</STEP_1>
  ...
  <STEP_n>Explain why the n-th intent is recommended</STEP_n>
</THOUGHT>

2. Intent sequence — A JSON array of tool-calling intents representing the recommendation. Each intent includes a tool name and associated parameters selected from a predefined intent library:

[
  {"工具名称": "tool_5", "起始位置": "当前位置", "空间范围": "附近", "tag": "美食"},
  {"工具名称": "tool_2", "起始位置": "当前位置", "终点位置": "家"},
  ...
]

The intent library includes 10 tool types covering scenarios such as ride-hailing, navigation, transit, POI recommendation, order reminders, weather queries, etc.

🚀 Quick Start

pip install -r requirements.txt
bash finetune.sh

📁 Project Structure

├── finetune.py                        # Finetuning script (WeightedLossTrainer + SyncEpochCallback)
├── finetune.sh                        # Launch script example
├── data_process/
│   ├── collate_fns.py                 # Data collator (progressive distillation)
│   └── data_loader.py                 # Data loading
├── utils.py                           # Utility functions and argument definitions
├── add_tokens/extended_cot_vocabs.json # CoT special token vocabulary
├── config/ds_z3_bf16.json             # DeepSpeed ZeRO-3 configuration
└── requirements.txt