abap-llm/docs/stage1-step-prompts.md

Stage 1: step prompts

Preparation (one time): copy stage1-training-task.md to ~/projects/abap-llm/harness/docs/stage1-training-task.md.

Use Sonnet. Before each prompt: /clear. Each prompt is one session.

---

A — Setup and data (A4H: no change needed)

Read CLAUDE.md and docs/stage1-training-task.md. Do only Step 0 and Step 1.
Create train/STATE.md: write what you did, the model id, the paths, the
report of Step 1, and the next step. Commit. Then stop.

B1 — Baseline, start (A4H must run)

Read CLAUDE.md, docs/stage1-training-task.md and train/STATE.md. Do Step 2.
Start the eval runs in the background with a log file and a macOS
notification at the end. Measure the valid loss of the base model in the
same background job, after the eval runs. Update train/STATE.md with the
log path and the job command. Then stop. Do not poll.

B2 — Baseline, results (after the notification)

Read CLAUDE.md and train/STATE.md. Read only the last 50 lines of the log.
Write runs/stage1/baseline.json. Update train/STATE.md with the results.
Commit. Show me a short summary. Then stop.

C — Training test (stop A4H first)

Read CLAUDE.md, docs/stage1-training-task.md and train/STATE.md. A4H is
stopped. Check that no other model is loaded. Do only the short test of
Step 3 (20 iterations). Report peak memory, time per iteration, and the
time estimate for the full run. If memory is not sufficient, give me
options and do not change the settings. Update train/STATE.md. Then stop.

D1 — Training, start (after my approval)

Read CLAUDE.md, docs/stage1-training-task.md and train/STATE.md. Start the
full training run of Step 3 in the background with a log file and a macOS
notification at the end. Use the settings from train/config.yaml. Update
train/STATE.md with the log path and the adapter path. Then stop. Do not
poll.

D2 — Training, results (after the notification)

Read CLAUDE.md and train/STATE.md. Read only the valid loss lines and the
last 30 lines of the log. Select the adapter with the lowest valid loss.
Update train/STATE.md. Commit. Show me the valid loss curve as a short
table. Then stop.

E1 — Measure again, start (start A4H first)

Read CLAUDE.md, docs/stage1-training-task.md and train/STATE.md. A4H runs.
Do Step 4: serve the base model with the selected adapter, connect the
harness llm agent to it (smallest change), and test it with 1 task. Then
start the same eval runs as in Step 2 and the valid loss measurement in the
background with a log file and a macOS notification at the end. Update
train/STATE.md. Commit. Then stop. Do not poll.

E2 — Report (after the notification)

Read CLAUDE.md, docs/stage1-training-task.md and train/STATE.md. Read only
the result files and the last 50 lines of the log. Write
runs/stage1/report.md as described in the task. Update train/STATE.md.
Commit. Show me the report. Then stop.