erhan/abap-llm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Devstral Small 2: serve.sh, baseline.py without thinking args, repair rate, activation error message fix, stop rule in chain (run base 21000)

Browse Source 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
This commit is contained in:
Kral
2026-10-04 08:57:12 +02:00
parent e84ea43a3d
commit 8d1db9c67c
5 changed files with 56 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
train/baseline_chain.sh
View File

@@ -1,20 +1,36 @@
#!/bin/sh
# Baseline on the 11-task subset (Kral decision 2026-10-03), detached. Replaces night_chain.sh.
# macOS notification after 2 tasks (time estimate) and at the end. Log: runs/stage1/baseline.log, chain log below.
# Baseline on the 11-task subset, Devstral Small 2 (2026-10-04), detached.
# Stop rule: after the first 4 tasks, if all 4 ended with "loop" and the repair rate is below 20 %, stop the run.
# Log: runs/stage1/baseline.log, chain log runs/stage1/baseline_chain.log
cd "$(dirname "$0")/.."
log() { echo "$(date '+%F %T') $*"; }
log "start baseline, 11 tasks, thinking off, max_tokens 16384, loop guard 3"
python3 train/baseline.py --label baseline --run-base 20500 > runs/stage1/baseline.log 2>&1 &
log "start baseline, 11 tasks, no thinking mode (Devstral), max_tokens 16384, loop guard 3"
python3 train/baseline.py --label baseline_devstral --run-base 21000 > runs/stage1/baseline.log 2>&1 &
BP=$!
log "baseline PID $BP"
NOTE=0
CHECKED=0
while kill -0 $BP 2>/dev/null; do
N=$(python3 -c "import json;print(len(json.load(open('runs/stage1/baseline.json'))['tasks']))" 2>/dev/null || echo 0)
if [ "$NOTE" = 0 ] && [ "$N" -ge 2 ]; then
osascript -e 'display notification "Baseline: first 2 tasks done, ask for the time estimate" with title "stage1"'
log "2 tasks done"; NOTE=1
if [ "$CHECKED" = 0 ]; then
R=$(python3 - <<'PY' 2>/dev/null
import json
t = json.load(open("runs/stage1/baseline_devstral.json"))["tasks"]
if len(t) < 4:
print("wait")
else:
f = list(t.values())[:4]
after = sum(x.get("pushes_after_error") or 0 for x in f)
ch = sum(x.get("pushes_changed_after_error") or 0 for x in f)
rate = ch / after if after else 0
loops = sum(x.get("end_reason") == "loop" for x in f)
print("stop" if loops == 4 and rate < 0.2 else "go", loops, round(rate, 3))
PY
)
case "$R" in
stop*) log "STOP RULE: $R"; kill $BP; osascript -e 'display notification "Stop rule: 4 loops, repair rate below 20 %" with title "stage1"'; exit 0 ;;
go*) log "first 4 tasks checked: $R"; CHECKED=1 ;;
esac
fi
sleep 60
done
log "baseline ended: $(python3 -c "import json;print(len(json.load(open('runs/stage1/baseline.json'))['tasks']))") tasks"
osascript -e 'display notification "Baseline (11 tasks) ended" with title "stage1"'
log "baseline ended: $(python3 -c "import json;print(len(json.load(open('runs/stage1/baseline_devstral.json'))['tasks']))") tasks"
osascript -e 'display notification "Baseline Devstral (11 tasks) ended" with title "stage1"'