File size: 3,516 Bytes
f9bc0c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
#!/usr/bin/env bash
# Build a single-file git bundle of everything the remote run needs, WITHOUT
# touching the (shared, "admin"-branch) repository that tracks ~/Desktop.
#
# What goes in (this mirror excludes junk; models included by default ~46 MB):
#   indra/            cloned inference package (Apache-2.0)
#   models/indra-ocr-v3.1.1/  ONNX + safetensors + meta (OpenRAIL-M)
#   tools/indra_eval/ verified eval harness + inspectors + synth generator
#   tools/exp/        THIS experiment code (train / export / exactness / A glue)
#   docs/             investigation + experiment plan + remote run sheet
#   results/smoke/    smoke records (tiny; gives the remote a live sanity ref)
#
# Usage: bash tools/exp/make_bundle.sh [--no-models] [--out FILE.bundle]
set -euo pipefail

REPO=$(cd "$(dirname "$0")/../.." && pwd)
STAGE=/tmp/opencode/indra_exp_stage
OUT=""
INCLUDE_MODELS=1
while [[ $# -gt 0 ]]; do
  case "$1" in
    --no-models) INCLUDE_MODELS=0; shift ;;
    --out) OUT="$2"; shift 2 ;;
    *) echo "unknown arg $1"; exit 1 ;;
  esac
done
OUT="${OUT:-$REPO/indra_nepali_exp.bundle}"

rm -rf "$STAGE"; mkdir -p "$STAGE"
mkdir -p "$STAGE/indra" "$STAGE/tools" "$STAGE/docs" "$STAGE/results/smoke"
rsync -a \
  --exclude '__pycache__' --exclude '*.pyc' --exclude '.ipynb_checkpoints' \
  --exclude '.git' --exclude '.github' --exclude 'indra.egg-info' \
  "$REPO/indra/" "$STAGE/indra/"
rsync -a --exclude '__pycache__' --exclude '*.pyc' "$REPO/tools/" "$STAGE/tools/"
rsync -a "$REPO/docs/" "$STAGE/docs/"
rsync -a "$REPO/logs/" "$STAGE/logs/"
rsync -a "$REPO/results/smoke/" "$STAGE/results/smoke/"
# top-level reference docs (tiny; the docx is the human-readable report)
for f in INDRA_OCR_NEPALI_COMPLETE_REPORT.docx INDRA_OCR_NEPALI_FEASIBILITY.md; do
  [[ -f "$REPO/$f" ]] && cp -a "$REPO/$f" "$STAGE/$f"
done
if [[ $INCLUDE_MODELS == 1 ]]; then
  mkdir -p "$STAGE/models"
  rsync -a "$REPO/models/" "$STAGE/models/"
fi

cat > "$STAGE/.gitignore" <<'EOF'
__pycache__/
*.pyc
.deps/
runs/
*.bundle
EOF

# a short top-level README for whoever lands on the far side
cat > "$STAGE/README_BUNDLE.md" <<'EOF'
# OCR_INDRA experiment bundle

Remote run companion for `docs/REMOTE_RUN_SHEET.md` (start there).
Layout: indra/ = cloned indra-ocr v0.3.2 (commit 3d773ae); models/indra-ocr-v3.1.1/
= Hub artifacts @ v3.1.1 (OpenRAIL-M — fine-tuning explicitly licensed §II ¶36);
tools/indra_eval = verified inspector/eval harness; tools/exp = experiment code;
INDRA_OCR_NEPALI_COMPLETE_REPORT.docx = full investigation report (read-only).

Setup: `bash tools/exp/remote_setup.sh` (deps + probe + offline verification gate),
then follow the run sheet.  `bash tools/exp/verify_bundle.sh` alone re-checks the
toolchain with no data: torch re-implementation → ONNX export → exactness vs the
shipped rec.onnx → rec-lines smoke CER.  Offline-verified 2026-10-20 on torch
2.14.1+cpu / onnxruntime 1.31.0 (export uses the legacy torch exporter on purpose).
Never copy datasets off the remote; only results files leave.
EOF

cd "$STAGE"
git init -q -b main
git -c user.name=ocr-indra-exp -c user.email=exp@local add -A
git -c user.name=ocr-indra-exp -c user.email=exp@local commit -qm "experiment bundle: indra eval + finetune toolkit"
git bundle create "$OUT" --all
git bundle verify "$OUT" >/dev/null 2>&1 && echo "bundle verified OK"

echo "bundle: $OUT ($(du -h "$OUT" | cut -f1))"
echo
echo "Transfer, then on the remote:"
echo "  git clone $OUT OCR_INDRA && cd OCR_INDRA && bash tools/exp/remote_setup.sh"