Usage Examples for convert2sgptArgs.py

Basic Usage (action only - default)

python convert2sgptArgs.py input_folder/ --output output.json

With All Tags (think, plan, memory)

python convert2sgptArgs.py input_folder/ --output output.json --think --plan --memory

Individual Tag Examples

Include thinking/reasoning sections

python convert2sgptArgs.py input_folder/ --output output.json --think

Include plan and step sections

python convert2sgptArgs.py input_folder/ --output output.json --plan

Include memory sections

python convert2sgptArgs.py input_folder/ --output output.json --memory

Combine multiple tags

python convert2sgptArgs.py input_folder/ --output output.json --think --plan
python convert2sgptArgs.py input_folder/ --output output.json --think --memory
python convert2sgptArgs.py input_folder/ --output output.json --plan --memory

Subsampling Trajectories with Probability p

You can randomly subsample trajectories (task folders) using the --p argument (default 1.0 means use all trajectories):

# Use 80% of trajectories on average
python convert2sgptArgs.py input_folder/ --output output.json --p 0.8

# Combine with tag options
python convert2sgptArgs.py input_folder/ --output output.json --think --plan --memory --p 0.5

The value of --p must be between 0 and 1. Each trajectory is independently included with probability p.

Sampling and Dropout

You can duplicate trajectories using the --sampling argument (default 1 means no duplication), and then apply dropout with --p:

# Sample each trajectory 6 times, then apply dropout with p=0.5
python convert2sgptArgs.py input_folder/ --output output.json --sampling 6 --p 0.5

# Combine with tag options
python convert2sgptArgs.py input_folder/ --output output.json --think --plan --sampling 3 --p 0.8

Order of operations:

1. First, each trajectory is duplicated sampling times (e.g., if --sampling 6, each trajectory becomes 6 copies)
2. Then, dropout p is applied to each duplicate independently (each copy has probability p of being kept)

Example: With 1 trajectory, --sampling 6, and --p 0.5:

• First: 6 copies are created
• Then: Each copy has a 50% chance of being kept
• Expected result: ~3 copies in the output

The value of --sampling must be >= 1.

What Each Tag Does

• --think: Includes <think> tags containing the agent's reasoning/thinking
• --plan: Includes <plan> and <step> tags containing planning information
• --memory: Includes <memory> tags containing memory updates
• Note: <action> tags are ALWAYS included regardless of flags

Example Output Structure

With all flags (--think --plan --memory), the output will contain:

{
  "system": "System prompt text...",
  "conversations": [
    {
      "from": "human",
      "value": "User prompt text..."
    },
    {
      "from": "gpt",
      "value": "<think>\nThinking content...\n</think>\n\n<plan>\nPlan content...\n</plan>\n\n<step>Step content...</step>\n\n<memory>\nMemory content...\n</memory>\n\n<action>\nAction content...\n</action>"
    }
  ]
}
```plan to include