Instructions

Primary Use Case

When a user requests to extract and save specific training metrics from experiment logs at regular step intervals (e.g., "every 100 steps") into a CSV file.

Core Workflow

1. Identify Target Experiment

• Input: User provides a target project/experiment (often via URL like wandb.ai/entity/project).
• Action: Query the project to list available runs/experiments.
• Decision Point: Determine which experiment to analyze based on user criteria (e.g., "shortest answers" → lowest response_length/mean).

2. Extract Metrics History

• Input: Selected experiment run identifier.
• Action: Query the run's history for the specific metrics requested by the user (e.g., actor/entropy_loss, response_length/clip_ratio, response_length/mean).
• Note: Use sampling queries to efficiently retrieve data at all steps.

3. Sample at Specified Intervals

• Input: Full history data and user-specified interval (e.g., "from step 0, at intervals of every 100 steps").
• Action: Filter the history to extract data points only at the requested steps (e.g., 0, 100, 200, ... up to the final step).
• Output: A structured list/dictionary of values for each target step.

4. Create CSV File

• Input: Sampled data with columns: step, metric1, metric2, ...
• Action: Format data as CSV with appropriate headers.
• Output: Write CSV file to the workspace with a descriptive filename (e.g., shortest_length_experiment.csv).

5. Provide Summary

• Action: Present a concise analysis summary to the user, highlighting:
  • Which experiment was selected and why.
  • Key observations from the extracted data (trends, min/max values).
  • Location and contents of the generated CSV file.

Key Tools & Patterns

• W&B Queries: Use wandb-query_wandb_tool for:
  • Project/run listing (ProjectInfo, GetRuns).
  • History keys inspection (RunHistoryKeys).
  • Sampled history data (RunHistorySampled with appropriate specs).
• File Operations: Use filesystem-write_file to create the CSV, and optionally filesystem-read_file to verify.
• Large Output Handling: When tool outputs are truncated, use local-view_overlong_tooloutput and related navigation/search tools to extract needed data.

Common User Phrases That Trigger This Skill

• "Record [metrics] into [filename].csv"
• "Save experiment data to a CSV file"
• "Extract data at intervals of every X steps"
• "Get the metrics from step 0, every 100 steps"
• "Analyze which experiment has the shortest/longest [metric]"

Error Handling & Edge Cases

• Missing Metrics: If a requested metric key doesn't exist in the run's history, inform the user and adjust the CSV columns accordingly.
• Insufficient Steps: If the run has fewer steps than the requested interval, sample all available steps and note the limitation.
• Large Datasets: For runs with many steps, use sampled history queries with appropriate max_items to avoid timeouts.