Skip to content

DANG-ai/SKILLER / skilllearnbench/skills/human_authored/temperature-simulation/glm-output

glm-output

Verified

Output-processing guidance for GLM tasks. Especially useful after glm-basics and glm-calibration when you need verifier-matching metrics from output.nc and a final /root/metrics.json.

GLM Output Guide

Overview

GLM produces NetCDF output containing simulated water temperature profiles. Processing this output requires understanding the coordinate system and matching with observations.

Suggested Skill Flow

This skill is usually the right final step after glm-basics and glm-calibration. As soon as calibration has found a configuration that appears to pass the task metrics, switch here, compute the final exact metrics, and write /root/metrics.json from this logic rather than from a separately improvised evaluator. If the exact final metrics pass here, end the task instead of returning to a broader calibration search.

Exact-Reporting Tasks

When a task requires an exact self-evaluation file, save /root/metrics.json with the task's required keys. For the deep-band calibration tasks in this benchmark family, the file must contain:

  • overall_rmse
  • annual_deep_rmse
  • summer_deep_rmse
  • overall_n_pairs
  • annual_deep_n_pairs
  • summer_deep_n_pairs

Do not use nearest-time approximations or vertical interpolation for the final reported metrics. Use the exact datetime/rounded-depth merge shown below, then compute any task-specific subsets from that merged table.

Output File

After running GLM, results are in output/output.nc:

VariableDescriptionShape
timeHours since simulation start(n_times,)
zHeight from lake bottom (not depth!)(n_times, n_layers, 1, 1)
tempWater temperature (°C)(n_times, n_layers, 1, 1)

Reading Output with Python

from netCDF4 import Dataset
import numpy as np
import pandas as pd
from datetime import datetime

nc = Dataset('output/output.nc', 'r')
time = nc.variables['time'][:]
z = nc.variables['z'][:]
temp = nc.variables['temp'][:]
nc.close()

Coordinate Conversion

Important: GLM z is height from lake bottom, not depth from surface.

LAKE_DEPTH = <lake_depth_from_nml>
depth_from_surface = LAKE_DEPTH - z

Complete Output Processing

from netCDF4 import Dataset
import numpy as np
import pandas as pd
from datetime import datetime

def read_glm_output(nc_path, lake_depth):
    nc = Dataset(nc_path, 'r')
    time = nc.variables['time'][:]
    z = nc.variables['z'][:]
    temp = nc.variables['temp'][:]
    start_date = datetime(2009, 1, 1, 12, 0, 0)

    records = []
    for t_idx in range(len(time)):
        hours = float(time[t_idx])
        date = pd.Timestamp(start_date) + pd.Timedelta(hours=hours)
        heights = z[t_idx, :, 0, 0]
        temps = temp[t_idx, :, 0, 0]

        for d_idx in range(len(heights)):
            h_val = heights[d_idx]
            t_val = temps[d_idx]
            if not np.ma.is_masked(h_val) and not np.ma.is_masked(t_val):
                depth = lake_depth - float(h_val)
                if 0 <= depth <= lake_depth:
                    records.append({
                        'datetime': date,
                        'depth': round(depth),
                        'temp_sim': float(t_val)
                    })
    nc.close()

    df = pd.DataFrame(records)
    df = df.groupby(['datetime', 'depth']).agg({'temp_sim': 'mean'}).reset_index()
    return df

Reading Observations

def read_observations(obs_path):
    df = pd.read_csv(obs_path)
    df['datetime'] = pd.to_datetime(df['datetime'])
    df['depth'] = df['depth'].round().astype(int)
    df = df.rename(columns={'temp': 'temp_obs'})
    return df[['datetime', 'depth', 'temp_obs']]

Calculating Task Metrics

def rmse_from_merged(merged):
    if len(merged) == 0:
        return 999.0
    errors = merged['temp_sim'] - merged['temp_obs']
    return float(np.sqrt(np.mean(errors ** 2)))

def calculate_task_metrics(sim_df, obs_df):
    merged = pd.merge(obs_df, sim_df, on=['datetime', 'depth'], how='inner')
    summer_mask = merged['datetime'].dt.month.isin([6, 7, 8, 9])
    annual_deep = merged[merged['depth'] >= 13]
    summer_deep = merged[summer_mask & (merged['depth'] >= 13)]

    return {
        'overall_rmse': rmse_from_merged(merged),
        'annual_deep_rmse': rmse_from_merged(annual_deep),
        'summer_deep_rmse': rmse_from_merged(summer_deep),
        'overall_n_pairs': int(len(merged)),
        'annual_deep_n_pairs': int(len(annual_deep)),
        'summer_deep_n_pairs': int(len(summer_deep)),
    }

Common Issues

IssueCauseSolution
RMSE very highWrong depth conversionUse lake_depth - z, not z directly
No matched observationsDatetime mismatchCheck datetime format consistency
Empty merged dataframeDepth rounding issuesRound depths to integers
Whole-lake RMSE looks fine but query still failsDeep subsets were not recomputed explicitlyCheck both depth >= 13 bands on the merged table

Best Practices

  • Check lake_depth in &init_profiles section of glm3.nml
  • Always convert z to depth from surface before comparing with observations
  • Round depths to integers for matching
  • Group by datetime and depth to handle duplicate records
  • Check the matched-pair counts for the full table, the annual deep band, and the summer deep band before writing metrics.json
  • Treat this as the final reporting handoff after calibration, not as another place to continue searching
  • Once the exact final metrics are written and pass, stop the task immediately
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall