Summarize Experiment

Generate a summary.md file capturing key metrics from a completed experiment. Think R's summary() for experiment results.

Your Task

Create a lightweight summary of experiment results:

1. Parse run status from experiment_summary.yaml
2. Extract final training loss from SLURM stdout
3. Extract accuracy from inspect-ai .eval files
4. Generate summary.md in experiment directory
5. Log the process in logs/summarize-experiment.log

Prerequisites

• experiment_summary.yaml exists
• At least some runs have completed (partial results acceptable)
• run-experiment has been executed (or manual SLURM jobs run)
• Conda environment activated - The parse_eval_log.py script requires inspect-ai. Activate the conda environment from claude.local.md before running extraction commands.

Workflow

1. Locate Experiment

Find the experiment directory:

• If in an experiment directory (contains experiment_summary.yaml): use current directory
• Otherwise: ask user for path

2. Parse Run Status

Read experiment_summary.yaml to identify runs:

From runs: section:

• name: Run identifier
• type: "fine-tuned" or "control"
• model: Model name
• parameters: Dict of hyperparameters (empty for control runs)

From evaluation.matrix: section:

• run: Run name
• tasks: List of evaluation task names
• epochs: List of epochs to evaluate (null for control runs)

Determine status by checking filesystem:

• Fine-tuning: Check for {output_base}/ck-out-{run_name}/ and SLURM outputs
• Evaluation: Check for {run_dir}/eval/logs/*.eval files

3. Extract Training Loss

For each COMPLETED fine-tuning run:

1. Find SLURM stdout in the output directory:
   • Parse experiment_summary.yaml "Output" section for output_dir_base
   • Look in: {output_dir_base}/ck-out-{run_name}/slurm-*.out
   • If multiple files, use most recent by modification time
2. Extract final loss using regex: (\d+)\|(\d+)\|Loss: ([0-9.]+)
   • Pattern matches: {epoch}|{step}|Loss: {value}
   • Take the LAST match to get final loss
   • The step number (group 2) from the last match is the total training steps
3. Record: run_name, final_loss, total_steps, epoch, step

Note: Training SLURM outputs are in the output directory, NOT the run directory.

If SLURM stdout missing:

• Log warning
• Record "N/A" for loss
• Continue with other runs

4. Extract Evaluation Accuracy

For each COMPLETED evaluation:

1. Find .eval files: {run_dir}/eval/logs/*.eval
2. For each .eval file, run:

   bash
   1python tools/inspect/parse_eval_log.py {path}

3. Parse JSON output for accuracy
4. Map to epoch using SLURM job names (see below)
5. For binary tasks, also run summary_binary.py to get balanced accuracy and F1
6. Record: run_name, task, epoch, accuracy, balanced_accuracy, f1, samples

Script output format:

json
1{
2  "status": "success",
3  "task": "capitalization",
4  "accuracy": 0.85,
5  "samples": 100,
6  "scorer": "exact_match",
7  "model": "..."
8}

Mapping Epochs via SLURM Job Names

The .eval files don't currently store epoch information directly. To reliably map each evaluation to its epoch:

1. Find SLURM output files in the eval directory: {run_dir}/eval/slurm-*.out
2. Extract job IDs from filenames (e.g., slurm-2773062.out → job ID 2773062)
3. Query job names via sacct:

   bash
   1sacct -j {job_ids} --format=JobID,JobName%50

4. Parse epoch from job name - scaffold-inspect names jobs like eval-{task}-{run}-ep{N}:
   • eval-general_eval-lowlr-ep0 → epoch 0
   • eval-general_eval-lowlr-ep9 → epoch 9
5. Extract accuracy from SLURM output:

   bash
   1grep -oP 'match/accuracy: \K[0-9.]+' slurm-{jobid}.out

Example workflow:

bash
1# Get job names for all eval jobs
2sacct -j 2773062,2773063,2773065 --format=JobID,JobName%50
3
4# Output shows epoch in job name:
5# 2773062  eval-general_eval-lowlr-ep0
6# 2773063  eval-general_eval-lowlr-ep1
7# 2773065  eval-general_eval-lowlr-ep2

This approach is reliable because:

• Job names are set by scaffold-inspect and include epoch info
• Works regardless of submission order or timing
• Survives job failures and resubmissions

If extraction fails:

• Script returns {"status": "error", "message": "..."}
• Log the error
• Record "ERROR" for accuracy
• Continue with other evaluations

Computing Balanced Accuracy and F1 (Binary Classification)

For binary classification tasks (0/1 targets), use summary_binary.py to compute additional metrics:

bash
1python tools/inspect/summary_binary.py {path_to_eval_file} --json

JSON output format:

json
1{
2  "status": "success",
3  "path": "/path/to/file.eval",
4  "samples": 100,
5  "accuracy": 0.85,
6  "balanced_accuracy": 0.83,
7  "f1": 0.82,
8  "precision_1": 0.80,
9  "recall_1": 0.84,
10  "recall_0": 0.82,
11  "confusion_matrix": {"tp": 42, "tn": 43, "fp": 7, "fn": 8, "other": 0}
12}

Why these metrics matter for imbalanced data:

• Balanced Accuracy = (Recall_0 + Recall_1) / 2 — not inflated by majority class
• F1 Score = harmonic mean of precision and recall — penalizes class imbalance

Note: For non-binary tasks, only accuracy is reported (Bal. Acc and F1 shown as "-").

5. Generate summary.md

Create {experiment_dir}/summary.md with the following structure:

markdown
1# Experiment Summary
2
3**Experiment:** `{experiment_name}` | **Generated:** {timestamp} | **Status:** {X}/{Y} complete
4
5## Run Status
6
7| Run | Type | Fine-tuning | Evaluation |
8|-----|------|-------------|------------|
9| rank4_lr1e-5 | Fine-tuned | COMPLETED | COMPLETED |
10| rank8_lr1e-5 | Fine-tuned | COMPLETED | COMPLETED |
11| base_model | Control | N/A | COMPLETED |
12
13## Training Results
14
15| Run | Final Loss | Total Steps | Epochs | Duration |
16|-----|------------|-------------|--------|----------|
17| rank4_lr1e-5 | 0.234 | 250 | 2 | 8m 15s |
18| rank8_lr1e-5 | 0.198 | 250 | 2 | 9m 02s |
19
20**Notes:**
21- Base model runs have no training loss (control)
22- Duration from SLURM elapsed time (if available)
23
24## Evaluation Results
25
26| Run | Task | Epoch | Accuracy | Bal. Acc | F1 | Samples |
27|-----|------|-------|----------|----------|------|---------|
28| rank4_lr1e-5 | capitalization | 0 | 0.85 | 0.83 | 0.82 | 100 |
29| rank4_lr1e-5 | capitalization | 1 | 0.88 | 0.86 | 0.85 | 100 |
30| rank8_lr1e-5 | capitalization | 0 | 0.82 | 0.80 | 0.78 | 100 |
31| rank8_lr1e-5 | capitalization | 1 | 0.91 | 0.89 | 0.88 | 100 |
32| base_model | capitalization | - | 0.45 | 0.50 | 0.31 | 100 |
33
34**Best performing:** rank8_lr1e-5 (epoch 1) with 89% balanced accuracy
35
36## Incomplete Runs
37
38| Run | Stage | Status | Notes |
39|-----|-------|--------|-------|
40| rank16_lr1e-5 | Fine-tuning | FAILED | Check slurm-12345.out |
41
42## Next Steps
43
441. View detailed evaluation results: `inspect view --port=$(get_free_port)`
452. Export raw data: `inspect log export {run_dir}/eval/logs/*.eval --format csv`
463. Full analysis: `analyze-experiment` (when available)
47
48---
49*Generated by summarize-experiment skill*

6. Create Log

Document the process in {experiment_dir}/logs/summarize-experiment.log.

See logging.md for action types and format.

Error Handling

If SLURM stdout missing

• Log warning with action type EXTRACT_LOSS
• Record "N/A" for loss in summary
• Continue with other runs

If .eval file cannot be parsed

• Log error with file path
• Record "ERROR" for accuracy in summary
• Continue with other evaluations

If all runs failed

• Generate summary noting all failures
• Include failure states in "Incomplete Runs" section
• Suggest troubleshooting steps

If partial results

• Generate summary with available data
• Clearly indicate which runs are missing in "Incomplete Runs" section
• Still identify best performing run from available data

Idempotency

Running summarize-experiment multiple times overwrites summary.md. This is intentional:

• Allows re-running after fixing failed runs
• Summary always reflects current state

Output Files

{experiment_dir}/
├── summary.md                    # Human-readable summary (new)
└── logs/
    └── summarize-experiment.log  # Process log (new)

Relationship to Other Skills

• After: run-experiment (or manual execution)
• Before: analyze-experiment (when available)
• Optional hook: run-experiment can invoke this at completion

Future Compatibility

When analyze-experiment is built, summarize-experiment can either:

• Remain as a quick summary option (text only, no plots)
• Be deprecated in favor of richer output
• Become a first stage that analyze-experiment builds upon