File size: 31,956 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 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 | """Build the complete Indra-OCR x Nepali investigation report as a .docx.
Assembles: cover + executive summary (fresh), the full contents of
docs/INDRA_INVESTIGATION.md, docs/INDRA_EXPERIMENTS.md, logs/investigation.md,
auto-generated smoke-results tables (results/smoke/*.json + GT records), and
an appendix (artifact inventory, dependency versions, reproduce commands).
Run: PYTHONPATH=/tmp/opencode/indra-lib python3 tools/indra_eval/make_docx.py
Out: INDRA_OCR_NEPALI_COMPLETE_REPORT.docx (workspace root)
"""
from __future__ import annotations
import json
import os
import re
import sys
import unicodedata
from pathlib import Path
ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, "/tmp/opencode/indra-lib")
from docx import Document # noqa: E402
from docx.enum.text import WD_ALIGN_PARAGRAPH # noqa: E402
from docx.oxml import OxmlElement # noqa: E402
from docx.oxml.ns import qn # noqa: E402
from docx.shared import Pt, RGBColor, Cm # noqa: E402
OUT = ROOT / "INDRA_OCR_NEPALI_COMPLETE_REPORT.docx"
# ---------------------------------------------------------------- md parser --
def parse_md(md: str):
"""Yield blocks: ('h',lvl,'..') | ('p','..') | ('li',kind,bool_nested,'..')
| ('code',[lines]) | ('table',[rows]) | ('hr',)"""
lines = md.splitlines()
i, out = 0, []
while i < len(lines):
ln = lines[i]
if ln.startswith("```"):
j, code = i + 1, []
while j < len(lines) and not lines[j].startswith("```"):
code.append(lines[j])
j += 1
out.append(("code", code))
i = j + 1
continue
if ln.strip() == "":
i += 1
continue
if ln.strip() == "---":
out.append(("hr",))
i += 1
continue
m = re.match(r"^(#{1,4})\s+(.*)$", ln)
if m:
out.append(("h", len(m.group(1)), m.group(2)))
i += 1
continue
if ln.lstrip().startswith("|"):
rows = []
while i < len(lines) and lines[i].lstrip().startswith("|"):
rows.append(lines[i].strip())
i += 1
out.append(("table", rows))
continue
m = re.match(r"^(\s*)([-*])\s+(.*)$", ln)
if m:
out.append(("li", "bullet", len(m.group(1)) >= 2, m.group(3)))
i += 1
continue
m = re.match(r"^(\s*)(\d+)\.\s+(.*)$", ln)
if m:
out.append(("li", "number", len(m.group(1)) >= 2, m.group(3)))
i += 1
continue
if ln.startswith(" ") and out and out[-1][0] in ("li", "p") and len(ln) - len(ln.lstrip()) >= 2 and ln.lstrip()[:2] != "- ":
# continuation of a list item / hanging paragraph
out.append(("cont", ln.strip()))
i += 1
continue
out.append(("p", ln))
i += 1
return out
# --------------------------------------------------------------- docx glue --
def cs_font(run):
"""Complex-script font so Devanagari renders consistently in Word."""
rpr = run._element.get_or_add_rPr()
rf = rpr.find(qn("w:rFonts"))
if rf is None:
rf = OxmlElement("w:rFonts")
rpr.append(rf)
rf.set(qn("w:cs"), "Nirmala UI")
def shade(par, fill="F4F4F4"):
ppr = par._p.get_or_add_pPr()
shd = OxmlElement("w:shd")
shd.set(qn("w:val"), "clear")
shd.set(qn("w:color"), "auto")
shd.set(qn("w:fill"), fill)
ppr.append(shd)
TOK = re.compile(r"(\*\*[^*]+\*\*|`[^`]+`|\*[^*]+\*)")
def add_runs(par, text, size=None, base_bold=False):
for tok in TOK.split(text):
if not tok:
continue
if tok.startswith("**") and tok.endswith("**") and len(tok) > 4:
r = par.add_run(tok[2:-2])
r.bold = True
cs_font(r)
elif tok.startswith("`") and tok.endswith("`") and len(tok) > 2:
r = par.add_run(tok[1:-1])
r.font.name = "Consolas"
r.font.size = Pt((size or 10.5) - 1)
cs_font(r)
elif tok.startswith("*") and tok.endswith("*") and len(tok) > 2:
r = par.add_run(tok[1:-1])
r.italic = True
cs_font(r)
else:
r = par.add_run(tok)
r.bold = base_bold or None
cs_font(r)
if size:
for r in par.runs:
pass
# normalize size on all runs of this paragraph if requested
if size:
for r in par.runs:
if r.font.size is None:
r.font.size = Pt(size)
def add_table(doc, rows, size=9):
vals = []
for r in rows:
if isinstance(r, str): # raw markdown row "| a | b |"
if re.match(r"^\|[\s:|-]+\|$", r.replace(" ", "")):
continue # separator row
cells = [c.strip() for c in r.strip().strip("|").split("|")]
else:
cells = [str(c) for c in r]
vals.append(cells)
if not vals:
return
ncol = max(len(v) for v in vals)
t = doc.add_table(rows=len(vals), cols=ncol)
t.style = "Table Grid"
for ri, row in enumerate(vals):
for ci in range(ncol):
cell = t.cell(ri, ci)
text = row[ci] if ci < len(row) else ""
par = cell.paragraphs[0]
add_runs(par, text, size=size, base_bold=(ri == 0))
par.paragraph_format.space_after = Pt(2)
def render_md(doc, md, first_h1_prefix=""):
"""Render markdown under the doc, # (first) -> Heading 1 (with prefix)."""
blocks = parse_md(md)
seen_h1 = False
for b in blocks:
kind = b[0]
if kind == "h":
lvl, text = b[1], b[2]
if lvl == 1:
if seen_h1:
continue # only one H1 per part
seen_h1 = True
text = first_h1_prefix + text
h = doc.add_heading(text, level=1) # used only for Part titles
for r in h.runs:
r.font.color.rgb = RGBColor(0x1F, 0x3B, 0x63)
continue
h = doc.add_heading("", level=lvl)
add_runs(h, text)
for r in h.runs:
r.font.color.rgb = RGBColor(0x22, 0x35, 0x52)
elif kind == "p":
par = doc.add_paragraph()
par.paragraph_format.space_after = Pt(4)
add_runs(par, b[1])
elif kind == "cont":
par = doc.paragraphs[-1]
par.add_run("\n" + b[1])
elif kind == "li":
_, lk, nested, text = b
style = "List Bullet" if lk == "bullet" else "List Number"
if nested:
style += " 2" if lk == "bullet" else ""
par = doc.add_paragraph(style=style)
par.paragraph_format.space_after = Pt(2)
add_runs(par, text)
elif kind == "code":
for ln in b[1]:
par = doc.add_paragraph()
par.paragraph_format.space_after = Pt(0)
par.paragraph_format.space_before = Pt(0)
r = par.add_run(ln if ln else " ")
r.font.name = "Consolas"
r.font.size = Pt(8.5)
cs_font(r)
shade(par)
doc.add_paragraph().paragraph_format.space_after = Pt(4)
elif kind == "table":
add_table(doc, b[1])
doc.add_paragraph().paragraph_format.space_after = Pt(4)
elif kind == "hr":
continue
def lev(a: str, b: str) -> int:
if a == b:
return 0
prev = list(range(len(b) + 1))
for i, ca in enumerate(a, 1):
cur = [i] + [0] * len(b)
for j, cb in enumerate(b, 1):
cur[j] = min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (ca != cb))
prev = cur
return prev[-1]
# ------------------------------------------------------------------ report --
def main():
doc = Document()
for sec in doc.sections:
sec.top_margin = sec.bottom_margin = Cm(2.0)
sec.left_margin = sec.right_margin = Cm(2.2)
st = doc.styles["Normal"]
st.font.name = "Calibri"
st.font.size = Pt(10.5)
cp = doc.core_properties
cp.title = "Indra OCR Γ Nepali (Devanagari) β Complete Engineering Investigation"
cp.author = "OCR_INDRA investigation"
cp.subject = "datalab-to/indra v0.3.2 / model revision v3.1.1 β evidence-based feasibility, adaptation and evaluation"
cp.comments = "All claims verified from source (commit 3d773ae) and model artifacts (ONNX + safetensors), not from README."
# ---- cover ----
doc.add_heading("Indra OCR Γ Nepali (Devanagari)", level=0)
p = doc.add_paragraph()
r = p.add_run("Complete Engineering Investigation Report")
r.font.size = Pt(16)
r.font.color.rgb = RGBColor(0x33, 0x33, 0x33)
meta = doc.add_paragraph()
add_runs(
meta,
"Package: `datalab-to/indra` v0.3.2 (commit `3d773ae`) Β· Model set: "
"HF revision `v3.1.1` Β· Workspace: `OCR_INDRA/` Β· Date: 2026-10-17 "
"(rev. 2026-10-20: fine-tune toolkit verified offline; bundle rebuilt)",
size=10,
)
doc.add_paragraph()
add_runs(
doc.add_paragraph(),
"Method: every architecture claim in this report is verified against the "
"actual artifacts β the cloned source (all 17 files of `indra/indra_ocr/`), "
"the 201-node ONNX graph of `rec.onnx`, the graph I/O of all 5 models, and "
"the shipped `model.safetensors` (108 F32 tensors) β never from README "
"marketing. Verification scripts are included and re-runnable "
"(Appendix B). Blocked items are marked [B]; verified items [V]; "
"refuted prior claims [X].",
)
doc.add_page_break()
# ---- Part I: executive summary ----
doc.add_heading("Part I β Executive summary", level=1)
h = doc.add_heading("Question", level=3)
add_runs(
doc.add_paragraph(),
"Can Datalab's Indra OCR be evaluated, adapted or fine-tuned for a "
"Nepali/Devanagari OCR use case β and could it beat the deployed "
"PaddleOCR PP-OCRv5 Mobile Devanagari pipeline? Decisions had to be "
"based on the actual source code and ONNX graphs, not the README.",
)
doc.add_heading("Seven key findings (all verified)", level=2)
add_runs(
doc.add_paragraph(style="List Number"),
"**The ~1 M-param recognizer is a padding-masked MobileNetV3-style 1-D "
"CNN with a flat CTC head β no RNN, no attention.** Exactly "
"**1,236,366 parameters** (safetensors β ONNX initializer cross-check). "
"Stem Conv 1β128 (kernel 32Γ4, stride 32Γ4) collapses the line height "
"entirely; then 10 identical residual blocks [depthwise 1Γ9 β expand "
"1Γ1 128β256 β HardSwish β SE (padding-aware) β project 1Γ1 256β128 "
"β +residual]; then a 1,678-class head.",
)
add_runs(
doc.add_paragraph(style="List Number"),
"**Time stride is 4, not 8** (the `WIDTH_MULTIPLE=8` comment in "
"`rec_preproc.py:24` is inaccurate as written). The graph computes "
"`lengths = ceil(widths/4)`, clipped to `W/4` (measured table, "
"e.g. widths 65 β 17). The graph outputs `ids [B,T]`, `conf [B,T]`, "
"`lengths [B]`; padded timesteps are NOT special-cased (head-on-zeros "
"β blank, conf β 0.78), so callers must truncate by `lengths` β as "
"`pipeline.py:297` does.",
)
add_runs(
doc.add_paragraph(style="List Number"),
"**Dynamic padding mask is functional, not cosmetic.** "
"`Cast_4 = (arange(W/4) < lengths).float()` zeroes the padded region "
"post-stem, after every block's residual (proven: final-trunk features "
"at padded t are exactly 0.0 even with ink present), and normalizes "
"every SE pooling over valid timesteps only.",
)
add_runs(
doc.add_paragraph(style="List Number"),
"**Vocabulary already covers 100% of Devanagari.** 1,677 tokens + "
"blank class 0 (head 1,678); all 128 assigned codepoints in "
"U+0900βU+097F verified present (letters, matras, nukta, danda ΰ₯€ ΰ₯₯, "
"Devanagari digits ΰ₯¦βΰ₯―, βΉ, ΰ₯). **No head surgery is needed for "
"Nepali** β the only adaptation question is training-data distribution.",
)
add_runs(
doc.add_paragraph(style="List Number"),
"**Fine-tuning is a reimplementation project, but legally and "
"architecturally allowed.** No training code exists in the repo "
"(grep-verified; `pyproject.toml` has no torch) and the artifacts "
"ship trainable weights ONLY for the recognizer (`model.safetensors`, "
"`rec/` prefix). det/order/orient/everything else exist as frozen "
"ONNX only. The weights license β AI PUBS Open RAIL-M (modified) β "
"explicitly permits fine-tuning (Β§II ΒΆ36); the inference code is "
"Apache-2.0.",
)
add_runs(
doc.add_paragraph(style="List Number"),
"**The evaluation harness is built and smoke-proven on real shaped "
"Devanagari.** Zero-shot full-pipeline run on a synthetic Nepali "
"page (raqm-shaped, OFL Noto Sans Devanagari): 10/10 lines detected, "
"reading order correct, micro CER **2.62%**; recognizer-only on the "
"identical GT crops: CER **3.49%**. Observed failure modes (nukta "
"ΰ€°ΰ€Όβΰ€° dropped, βΉβΰ€°, ΰ₯βΰ€, conjunct matras `ΰ€Έΰ€ΰ₯ΰ€ΰ₯ΰ€€βΰ€Έΰ€ΰ₯ΰ€ΰ₯ΰ€€`) are exactly "
"the distribution gaps that fine-tuning (Experiment B) targets.",
)
add_runs(
doc.add_paragraph(style="List Number"),
"**The fine-tune toolkit is built and verified offline (2026-10-20).** "
"`tools/exp/` reconstructs the recognizer in PyTorch from the shipped "
"safetensors, fine-tunes with CTC, and re-exports ONNX: the zero-shot "
"export matches the shipped `rec.onnx` with 0 id mismatches and max "
"|Ξconf| 7.1e-6 (gate 1e-3); a mechanics training run (loss 8.1β1.65, "
"val CER 0.97β0.58 on an overfit probe) passes the post-train gate "
"(torch `.pt` vs its own ONNX, Ξconf 1.3e-5) and `run_rec.py` reproduces "
"the recorded smoke CER (3.49 %) bit-for-bit. `verify_bundle.sh` pins "
"all of this in one data-free command, and the rebuilt "
"`indra_nepali_exp.bundle` carries it to the remote.",
)
doc.add_heading("Verdict", level=2)
add_runs(
doc.add_paragraph(),
"**Zero-shot evaluation (Experiment A) is ready today** β the adapter "
"(`tools/indra_eval/eval_indra.py`) runs Indra end-to-end on CPU and "
"scores CER/WER, and the Experiment B toolchain is verified offline "
"(`tools/exp/verify_bundle.sh`); only the frozen Nepali eval sets on the "
"remote machine are missing (`khalti-t4` β SSH still rejected for the "
"presented key on 2026-10-20: authorize it or hand over the bundle). "
"**The recognizer is the only component worth adapting**, via the "
"shipped safetensors init; a re-export must reproduce the padding "
"mask + per-block re-zeroing + SE valid-means + the "
"`ids/conf/lengths` decode tail. Detector/order fine-tuning is "
"out of scope until the zero-shot full-pipeline run shows detection "
"as the bottleneck (Experiment A). Recommended order: **A β D (hybrid "
"PP-det + Indra-rec, zero training) β B (rec fine-tune) β E (int8 "
"deploy)**.",
)
doc.add_page_break()
# ---- Part II: full investigation doc ----
doc.add_heading("Part II β Architecture investigation (full evidence)", level=1)
md = (ROOT / "docs" / "INDRA_INVESTIGATION.md").read_text(encoding="utf-8")
# drop the md's own title + status legend line (already on cover)
parts = re.split(r"^## 0\. Verdict", md, maxsplit=1, flags=re.M)
md = "## 0. Verdict" + parts[1]
render_md(doc, md, first_h1_prefix="")
doc.add_page_break()
# ---- Part III: experiments doc ----
doc.add_heading("Part III β Experiment plan (AβE) and governance", level=1)
md = (ROOT / "docs" / "INDRA_EXPERIMENTS.md").read_text(encoding="utf-8")
render_md(doc, md, first_h1_prefix="")
doc.add_page_break()
# ---- Part IV: log ----
doc.add_heading("Part IV β Investigation log", level=1)
md = (ROOT / "logs" / "investigation.md").read_text(encoding="utf-8")
render_md(doc, md, first_h1_prefix="")
doc.add_page_break()
# ---- Part V: smoke results (generated from artifacts) ----
doc.add_heading("Part V β Smoke-test results (generated from result files)", level=1)
gt = json.loads((ROOT / "results/smoke/synth_nepali_record.jsonl").read_text(encoding="utf-8").splitlines()[0])
rec_only = json.loads((ROOT / "results/smoke/rec_only.jsonl").read_text().splitlines()[0])
full = json.loads((ROOT / "results/smoke/full_run.jsonl").read_text().splitlines()[0])
summary = {
"mode": {
"rec-only": rec_only,
"full pipeline": full,
}
}
doc.add_heading("V.1 Setup", level=2)
add_runs(
doc.add_paragraph(),
"Synthetic A4-ratio page (1240Γ1754 px), 10 Nepali lines rendered at "
"32 px with raqm shaping (conjuncts, nukta, digits, ΰ₯, βΉ, danda). "
"Ground truth is exact by construction. Weights: Indra v3.1.1 "
"official, CPU (onnxruntime 1.31.0). **Numbers are an environment "
"smoke test, not a benchmark** β the frozen remote eval sets are "
"needed for the real comparison (Experiment A).",
)
doc.add_heading("V.2 Headline numbers", level=2)
rows = []
for mode, rec in summary["mode"].items():
cer = lev_text_stats(rec, gt)
rows.append([mode, str(len(rec["lines"])), f"{rec['mean_conf']:.3f}", f"{cer['cer']:.4f}", f"{cer['wer']:.4f}"])
add_table(
doc,
[["mode", "lines found", "mean conf", "micro CER", "micro WER"]]
+ rows,
)
doc.add_paragraph()
doc.add_heading("V.3 Per-line comparison (recognizer-only, GT crops)", level=2)
rows = []
for i, (gl, pl) in enumerate(zip(gt["lines"], rec_only["lines"])):
g = unicodedata.normalize("NFC", gl["text"])
p = unicodedata.normalize("NFC", pl["text"])
d = lev(g, p)
mark = "β" if g == p else f"{d} edits"
rows.append([str(i), g, p, f"{pl['conf']:.3f}", mark])
add_table(doc, [["#", "ground truth", "prediction", "conf", "Ξ"], *rows], size=9)
doc.add_paragraph()
doc.add_heading("V.4 Per-line comparison (full pipeline, detector crops)", level=2)
rows = []
for i, (gl, pl) in enumerate(zip(gt["lines"], full["lines"])):
g = unicodedata.normalize("NFC", gl["text"])
p = unicodedata.normalize("NFC", pl["text"])
d = lev(g, p)
mark = "β" if g == p else f"{d} edits"
rows.append([str(i), g, p, f"{pl['conf']:.3f}", mark])
add_table(doc, [["#", "ground truth", "prediction", "conf", "Ξ"], *rows], size=9)
doc.add_paragraph()
doc.add_heading("V.5 Interpretation", level=2)
add_runs(
doc.add_paragraph(),
"Both runs read the page almost perfectly; the remaining errors are "
"print-size artifacts (32 px) of exactly the classes a Nepali corpus "
"would reinforce: the nukta dot (ΰ€°ΰ€Ό), the rarely-trained βΉ and ΰ₯, and "
"conjunct vmatras. The rec-only vs full discrepancy on line 2 "
"(ΰ₯ β ΰ€ vs S) and line 3/5/7 flips demonstrates crop-offset "
"sensitivity β which is why the fair design compares recognizers on "
"**identical crops** (rec-lines mode) and pipelines on page text. "
"Timings on this WSL box are not representative (CPU/IO-starved; det "
"β 18.6 min on one page) β benchmark on the T4.",
)
doc.add_page_break()
# ---- Part VI: appendix ----
doc.add_heading("Part VI β Appendix", level=1)
doc.add_heading("A. Artifact inventory", level=2)
rows = []
for rel, why in [
("docs/INDRA_INVESTIGATION.md", "architecture + evidence + [U] closure"),
("docs/INDRA_EXPERIMENTS.md", "experiments AβE, leakage, licensing, decision tree"),
("logs/investigation.md", "command/evidence log"),
("tools/indra_eval/inspect_vocab.py", "vocab + Devanagari coverage verification"),
("tools/indra_eval/inspect_models.py", "graph I/O + param-count cross-check"),
("tools/indra_eval/inspect_rec_nodes.py", "full rec.onnx node dump + numeric run"),
("tools/indra_eval/inspect_rec_mask.py", "mask semantics + lengths table proof"),
("tools/indra_eval/make_synth_page.py", "shaped Devanagari test page + GT"),
("tools/indra_eval/eval_indra.py", "predict (full / rec-lines) + CER scorer"),
("tools/exp/torch_rec.py", "PyTorch re-implementation of rec.onnx + weight map"),
("tools/exp/train_rec.py", "CTC fine-tune loop (safetensors init, width-bucketed batches)"),
("tools/exp/export_rec_onnx.py", "torch -> ONNX with the exact ids/conf/lengths contract"),
("tools/exp/exactness_check.py", "gate: export vs rec.onnx / torch (ids, lengths, conf)"),
("tools/exp/run_rec.py", "batch rec-lines runner (ONNX or --torch backends)"),
("tools/exp/verify_bundle.sh", "one-command offline verification gate (no data needed)"),
("tools/exp/e_int8_quantize.py", "Experiment E: static (QDQ) / dynamic int8 quantization"),
("tools/exp/run_a.sh", "Experiment A orchestration (rec-lines + GT records + CER)"),
("tools/exp/run_b.sh", "Experiment B orchestration (train -> export -> gate -> re-score)"),
("tools/exp/data_prep.py", "corpus -> manifests, page-hash split, leakage exclude"),
("tools/exp/a_export_gt_boxes.py", "GT boxes adapters (validate/normalize/merge)"),
("tools/exp/a_ppocr_baseline.py", "identical-crop baseline materials (engine = remote plug)"),
("tools/exp/remote_setup.sh", "remote deps + probe + verification gate"),
("tools/exp/make_bundle.sh", "builds indra_nepali_exp.bundle for the remote"),
("tools/indra_eval/testdata/NotoSansDevanagari.ttf", "OFL font (test rendering only)"),
("models/indra-ocr-v3.1.1/", "downloaded official weights (16 files, ~46 MB)"),
("results/smoke/", "smoke artifacts (pages, records, CER JSON)"),
("results/verify/", "offline verification evidence (exactness + smoke re-run)"),
]:
pth = ROOT / rel
n = (
sum(f.stat().st_size for f in pth.rglob("*") if f.is_file())
if pth.is_dir()
else (pth.stat().st_size if pth.exists() else 0)
)
rows.append([rel, why, f"{n / 1024:.0f} KiB" if n else "β"])
add_table(doc, [["path", "purpose", "size"], *rows], size=9)
doc.add_paragraph()
doc.add_heading("B. Reproduce everything", level=2)
code = '\n'.join(
[
"cd /mnt/c/Users/Dell/Desktop/OCR_INDRA # or wherever this workspace is",
"PY=/tmp/opencode/indra-lib # pip --target deps",
"export PYTHONPATH=\"$PWD/indra:$PY\"",
"",
"python3 tools/indra_eval/inspect_vocab.py # vocab + Devanagari coverage",
"python3 tools/indra_eval/inspect_models.py # graph I/O + param count",
"python3 tools/indra_eval/inspect_rec_nodes.py # full node dump + numeric run",
"python3 tools/indra_eval/inspect_rec_mask.py # mask semantics + lengths table",
"",
"python3 tools/indra_eval/make_synth_page.py results/smoke # synth page + GT",
"python3 tools/indra_eval/eval_indra.py predict results/smoke/synth_nepali_page.png \\",
" --boxes results/smoke/synth_nepali_gt.jsonl --jsonl results/smoke/rec_only.jsonl",
"python3 tools/indra_eval/eval_indra.py predict results/smoke/synth_nepali_page.png \\",
" --jsonl results/smoke/full_run.jsonl --print-text",
"python3 tools/indra_eval/eval_indra.py cer results/smoke/synth_nepali_record.jsonl \\",
" results/smoke/rec_only.jsonl --detail results/smoke/cer_rec_only.json --rec-only",
"python3 tools/indra_eval/eval_indra.py cer results/smoke/synth_nepali_record.jsonl \\",
" results/smoke/full_run.jsonl --detail results/smoke/cer_full.json",
"",
"# fine-tune toolkit: one-command offline verification (no data needed)",
"bash tools/exp/verify_bundle.sh # -> results/verify/summary.json {pass: true}",
"",
"python3 tools/indra_eval/make_docx.py # regenerate THIS report",
"",
"# remote (after SSH/bundle access): docs/REMOTE_RUN_SHEET.md",
"# bash tools/exp/remote_setup.sh # deps + probe + same verify gate",
"# bash tools/exp/run_a.sh # zero-shot A on the frozen sets",
"# bash tools/exp/run_b.sh # fine-tune B (train -> export -> gate -> re-score)",
]
)
for ln in code.split("\n"):
par = doc.add_paragraph()
par.paragraph_format.space_after = Pt(0)
r = par.add_run(ln if ln else " ")
r.font.name = "Consolas"
r.font.size = Pt(8.5)
cs_font(r)
shade(par)
doc.add_paragraph()
doc.add_heading("C. Environment & dependencies", level=2)
add_table(
doc,
[
["component", "version / note"],
["OS / box", "Linux (WSL-backed /mnt/c), 8 cores, 3 GB RAM β timings NOT representative"],
["Python", "3.12.3 system; deps via pip --target /tmp/opencode/indra-lib"],
["torch", "2.14.1+cpu β export forced to the legacy TorchScript tracer (see run sheet Β§0)"],
["onnxruntime", "1.31.0 (CPU provider)"],
["Pillow", "13.0.0 β raqm available (Devanagari shaping verified)"],
["opencv / onnx / pypdfium2 / hf-hub / safetensors / numpy", "installed in the same target dir"],
["python-docx", "1.2.0 (this report's generator)"],
["Remote machine", "khalti-t4 (202.51.0.7) β SSH key still rejected 2026-10-20 (port 22 open); all dataset-side work blocked"],
],
)
doc.add_paragraph()
doc.add_heading("D. Licensing (verified from files)", level=2)
add_table(
doc,
[
["item", "license", "fine-tune", "deploy / notes"],
["Indra v3.1.1 model weights", "AI PUBS Open RAIL-M (modified) v0.1",
"**explicitly allowed** (Β§II ΒΆ36); Attachment A use restrictions attach",
"redistribution keeps license + usage notice; derivatives inherit it"],
["indra inference code", "Apache-2.0", "β", "standard reuse rights"],
["PaddleOCR (baseline, remote)", "Apache-2.0", "already fine-tuned", "deployed today"],
["Noto Sans Devanagari (test font)", "SIL OFL 1.1", "β", "rendering only"],
["Nepali datasets / frozen eval sets", "docs/DATA_LICENSING.md [B]",
"[B] review pending SSH", "no datasets leave the remote machine"],
],
)
doc.add_page_break()
# ---- Part VII: toolkit status + remote readiness (2026-10-20) ----
doc.add_heading("Part VII β Fine-tune toolkit & remote readiness (2026-10-20)", level=1)
add_runs(
doc.add_paragraph(),
"The Experiment A/B toolchain in `tools/exp/` was completed and verified "
"offline this revision; the rebuilt single-file bundle "
"`indra_nepali_exp.bundle` carries it to `khalti-t4`. No dataset was "
"touched (none is reachable).",
)
doc.add_heading("Verified offline", level=2)
add_table(
doc,
[
["check", "command", "result"],
["bundle integrity", "`git bundle verify` + fresh clone + byte-compare", "OK β complete history, 16 model files byte-identical"],
["param invariant", "`train_rec.py --selfcheck`", "1,236,366 tensors (parameters 1,236,110 + BN buffers 256)"],
["zero-shot export parity", "`export_rec_onnx.py --safetensors` vs `rec.onnx`", "**pass** β 0 id mismatches, 0 lengths mismatches, max |Ξconf| 7.1e-6"],
["training mechanics", "`train_rec.py` 120 steps, overfit probe", "loss 8.1 β 1.65; val CER 0.97 β 0.58; checkpoints + `best.safetensors` written"],
["post-train gate", "`exactness_check.py --against torch`", "**pass** β 0 id mismatches, Ξconf 1.3e-5"],
["backend parity", "`run_rec.py` ONNX vs `--torch`", "identical texts on the smoke page"],
["A orchestration", "`run_a.sh` on a local stand-in set", "GT records β rec-lines β CER 0.0349 (matches smoke exactly)"],
["B orchestration", "`run_b.sh` (train β export β gate β re-score)", "complete; findings written under `results/`"],
["data adapters", "`a_export_gt_boxes.py`, `a_ppocr_baseline.py`, `data_prep.py`", "all subcommands exercised OK"],
],
size=9,
)
doc.add_paragraph()
doc.add_heading("Fixed this revision (each would have broken the remote run)", level=2)
add_runs(
doc.add_paragraph(style="List Bullet"),
"torch β₯ 2.6 `torch.onnx.export` defaults to the torch.export exporter "
"(needs `onnxscript`); the export now pins the verified legacy tracer "
"(`dynamo=False`) with a fallback for older torch.",
)
add_runs(
doc.add_paragraph(style="List Bullet"),
"`run_a.sh` / `run_b.sh` crashed on unset `PYTHONPATH` under `set -u`.",
)
add_runs(
doc.add_paragraph(style="List Bullet"),
"`a_ppocr_baseline.py` rejected the `\"file\"` GT schema and did not "
"resolve page paths relative to the boxes file.",
)
add_runs(
doc.add_paragraph(style="List Bullet"),
"`remote_setup.sh` wrote `probe.json` relative to the caller's cwd and "
"ignored combined flags; it now also runs the full verification gate.",
)
add_runs(
doc.add_paragraph(style="List Bullet"),
"`data_prep.py` hash-split could leave val empty on small corpora.",
)
doc.add_heading("Remote status", level=2)
add_runs(
doc.add_paragraph(),
"SSH to `khalti-t4` (202.51.0.7) is still rejected for key "
"`khalti-t4-ocr-team` (`Permission denied (publickey,password)`; port 22 "
"open, 2026-10-20). To proceed: authorize the key's public half on the "
"box (or transfer the bundle to someone with access), then follow "
"`docs/REMOTE_RUN_SHEET.md` β `remote_setup.sh` (deps + probe + gate) β "
"probe inventory (Β§1) β `run_a.sh` β `run_b.sh`. Only `results/`, "
"`runs/rec_ft/*` and `results/probe.json` travel back; datasets never "
"leave the machine.",
)
doc.save(OUT)
print(f"written: {OUT} ({OUT.stat().st_size / 1024:.0f} KiB)")
def lev_text_stats(rec, gt):
gt_text = [unicodedata.normalize("NFC", l["text"]) for l in gt["lines"]]
pr_text = [unicodedata.normalize("NFC", l["text"]) for l in rec["lines"]]
ce = sum(lev(a, b) for a, b in zip(gt_text, pr_text))
cn = sum(len(a) for a in gt_text)
gw = " ".join(gt_text).split()
pw = " ".join(pr_text).split()
prev = list(range(len(pw) + 1))
for i, w in enumerate(gw, 1):
cur = [i] + [0] * len(pw)
for j, v in enumerate(pw, 1):
cur[j] = min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (w != v))
prev = cur
return {"cer": ce / max(1, cn), "wer": prev[-1] / max(1, len(gw))}
if __name__ == "__main__":
main()
|