View Hawk Eval Results

When the user wants to analyze evaluation results, use these hawk CLI commands:

1. List Eval Sets

You can list all eval sets if the user do not know the eval set ID:

hawk list eval-sets

Shows: eval set ID, creation date, creator.

You can increase the limit of results returned by --limit N.

hawk list eval-sets --limit 50

Or you can search for a specific eval set by using --search QUERY.

hawk list eval-sets --search pico

2. List Evaluations

With an eval set ID, you can list all evaluations in the eval-set:

hawk list evals [EVAL_SET_ID]

Shows: task name, model, status (success/error/cancelled), and sample counts.

3. List Samples

Or you can list individual samples and their scores:

hawk list samples [EVAL_SET_ID] [--eval FILE] [--limit N]

4. Download Transcript

To get the full conversation for a specific sample:

hawk transcript <UUID>

The transcript includes full conversation with tool calls, scores, and metadata.

To get even more details, you can get the raw data by using --raw:

hawk transcript <UUID> --raw

Batch Transcript Download

You can also download all transcripts for an entire eval set:

# Fetch all samples in an eval set
hawk transcripts <EVAL_SET_ID>

# Write to individual files in a directory
hawk transcripts <EVAL_SET_ID> --output-dir ./transcripts

# Limit number of samples
hawk transcripts <EVAL_SET_ID> --limit 10

# Raw JSON output (one JSON per line to stdout, or .json files with --output-dir)
hawk transcripts <EVAL_SET_ID> --raw

Known Limitations

• hawk list samples has a max --limit of 500 (API returns 422 for higher values)
• hawk transcript and hawk transcripts time out on large eval files (100MB+), common with side-task evals that have thousands of samples × multiple epochs
• hawk list samples does not index score values — the score_value field is often None

Prefer the Data Warehouse for Bulk Analysis

For querying sample-level data across eval sets (scores, limits, errors, token counts), use the warehouse-query skill instead of downloading eval files from S3 via inspect_ai.log.read_eval_log(). The warehouse has eval, sample, score, and message tables. A SQL query takes seconds vs minutes/hours for large eval files.

Example — find all samples that hit the working limit:

SELECT s.id AS sample_id, s.epoch, e.eval_set_id, e.model, e.task_args
FROM sample s
JOIN eval e ON s.eval_pk = e.pk
WHERE e.eval_set_id = 'eval-set-xxx'
  AND s."limit" = 'working';

Use hawk transcript <uuid> only when you need the full conversation transcript for a specific sample.

Workflow

1. Run hawk list eval-sets to see available eval sets 2a. Run hawk list evals <EVAL_SET_ID> to see available evaluations 2b. or run hawk list samples <EVAL_SET_ID> to find samples of interest (max 500 per request)
2. For bulk sample-level analysis (scores, limits, errors), use warehouse-query skill with SQL
3. Run hawk transcript <uuid> only for full conversation details on individual samples

API Environments

Production (https://api.inspect-ai.internal.metr.org) is used by default. Set HAWK_API_URL only when targeting non-production environments:

Example:

HAWK_API_URL=https://api.inspect-ai.staging.metr-dev.org hawk list eval_sets