Export evaluation data

• Pulling metrics into spreadsheets or notebooks for custom analysis and visualization.
• Feeding evaluation results into CI/CD pipelines to gate deployments.
• Sharing results with stakeholders who don’t have W&B seats, through BI tools like Looker or internal dashboards.
• Building automated reporting pipelines that aggregate scores across projects.

​

API endpoints used

• GET /v2/{entity}/{project}/evaluation_runs: Lists evaluation runs in a project, with optional filters by evaluation reference, model reference, or run ID.
• GET /v2/{entity}/{project}/evaluation_runs/{evaluation_run_id}: Reads a single evaluation run to retrieve its model, evaluation reference, status, timestamps, and summary.
• POST /v2/{entity}/{project}/eval_results/query: Retrieves grouped evaluation result rows for one or more evaluations. Returns per-row trials with model output, scores, and optionally resolved dataset row inputs. Also returns aggregated scorer statistics when requested.
• GET /v2/{entity}/{project}/predictions/{prediction_id}: Reads an individual prediction with its inputs, output, and model reference.

​

Prerequisites

• Python 3.7 or later.
• The requests library. Install it with pip install requests.
• A W&B API key, set as the WANDB_API_KEY environment variable. Get your key at wandb.ai/settings.

​

Set up authentication

import json
import os

import requests

TRACE_BASE = "https://trace.wandb.ai"
AUTH = ("api", os.environ["WANDB_API_KEY"])

entity = "my-team"
project = "my-project"

​

List evaluation runs

resp = requests.get(
    f"{TRACE_BASE}/v2/{entity}/{project}/evaluation_runs",
    auth=AUTH,
)
runs = [json.loads(line) for line in resp.text.strip().splitlines()]

for run in runs:
    print(run["evaluation_run_id"], run.get("status"))

​

Read a single evaluation run

eval_run_id = "<evaluation-run-id>"

resp = requests.get(
    f"{TRACE_BASE}/v2/{entity}/{project}/evaluation_runs/{eval_run_id}",
    auth=AUTH,
)
eval_run = resp.json()
print(eval_run["evaluation_run_id"], eval_run.get("status"), eval_run.get("model"))

​

Get predictions and scores

eval_run_id = "<evaluation-run-id>"

resp = requests.post(
    f"{TRACE_BASE}/v2/{entity}/{project}/eval_results/query",
    json={
        "evaluation_run_ids": [eval_run_id],
        "include_rows": True,
        "include_raw_data_rows": True,
        "resolve_row_refs": True,
    },
    auth=AUTH,
)
results = resp.json()

for row in results["rows"]:
    inputs = row.get("raw_data_row")
    for ev in row.get("evaluations", []):
        for trial in ev.get("trials", []):
            output = trial.get("model_output")
            scores = trial.get("scores", {})
            print("Input:", inputs)
            print("Output:", output)
            print("Scores:", scores)

​

Get aggregated scores

resp = requests.post(
    f"{TRACE_BASE}/v2/{entity}/{project}/eval_results/query",
    json={
        "evaluation_run_ids": [eval_run_id],
        "include_summary": True,
        "include_rows": False,
    },
    auth=AUTH,
)
results = resp.json()

for ev in results["summary"]["evaluations"]:
    for stat in ev["scorer_stats"]:
        print(stat["scorer_key"], stat.get("value_type"), stat.get("pass_rate") or stat.get("numeric_mean"))

​

Read a single prediction

prediction_id = "<predict-call-id>"

resp = requests.get(
    f"{TRACE_BASE}/v2/{entity}/{project}/predictions/{prediction_id}",
    auth=AUTH,
)
prediction = resp.json()
print(prediction)

​

How to use row digests

• Cross-evaluation comparison: When you run two different models against the same dataset, rows with the same digest represent the same input. You can join on row_digest to compare how different models performed on the exact same task.
• Deduplication: If the same task appears in multiple evaluation suites, the digest lets you identify it.
• Reproducibility: The digest is content-addressable, so if someone modifies a dataset row (changes the instruction text, rubric, or other fields), it gets a new digest. You can verify whether two evaluation runs used identical inputs or slightly different versions.