trl-sft - Experiment Metrics

This dataset contains training metrics logged with nvizio.

Structure

runs/
  run_name_1/
    metrics.parquet
    config.json
  run_name_2/
    metrics.parquet
    config.json
  ...

Usage

from nvizio.hub import list_runs_in_hub, fetch_run_from_hub

# List all runs
runs = list_runs_in_hub("kashif/nvizio-test-metrics")
print(runs)  # ['run_name_1', 'run_name_2', ...]

# Fetch a specific run
data = fetch_run_from_hub("kashif/nvizio-test-metrics", run_name="run_name_1")
print(data["metrics"].keys())  # ['train/loss', 'train/accuracy', ...]

Query with DuckDB

import duckdb

# Query all runs
conn = duckdb.connect()
df = conn.execute("""
    SELECT * FROM 'hf://datasets/kashif/nvizio-test-metrics/runs/*/metrics.parquet'
    WHERE tag LIKE 'train/%'
""").df()

# Query a specific run
df = conn.execute("""
    SELECT * FROM 'hf://datasets/kashif/nvizio-test-metrics/runs/my_run/metrics.parquet'
""").df()