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

Stage 1 step 1: prepare.py, real corpus numbers (SAP-samples/abap-cheat-sheets)

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
2026-10-03 21:58:04 +02:00
parent f0593933f9
commit e083c9ca13
6 changed files with 593 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
.gitignore vendored
View File

@@ -4,3 +4,4 @@ runs/
__pycache__/
.DS_Store
train/.venv/
train/data/*.jsonl

8
docs/stage1-training-task.md
View File

@@ -11,10 +11,12 @@ Read CLAUDE.md first. This task adds stage 1 of the training plan:
Stage 2 (SFT on teacher trajectories) is not part of this task.
Input: `~/projects/abap-llm/corpus/corpus.jsonl` (about 1,765 documents,
about 1.4M tokens, estimate). One record per line:
Input: `~/projects/abap-llm/corpus/corpus.jsonl`. Source (since 2026-10-03):
SAP-samples/abap-cheat-sheets, Apache-2.0 (`corpus/out/ATTRIBUTION.md`).
Real numbers from Step 1 (`train/data/report.md`): 370 documents, 2,806,501 tokens (Qwen tokenizer;
the chars/4 estimate is 2.64M). The old numbers (1,765 documents, 1.4M tokens) are outdated. One record per line:
`{"text", "objects", "types", "language_version", "package_path",
"grouped", "obsolete", "tokens"}`. Token counts are estimates (chars / 4).
"grouped", "obsolete", "tokens"}`. The `tokens` field is an estimate (chars / 4); there is also a field `source`.
## Rules

8
train/STATE.md
View File

@@ -15,6 +15,11 @@ Task: `docs/stage1-training-task.md`. Settings and weights: `train/README.md`. U
- Runner: `train/baseline.py --label <label>` (one task at a time, results `runs/stage1/<label>.json`,
run directories `runs/stage1/<label>/`).
- Step 1 done (2026-10-03 22:00): `train/prepare.py`, report `train/data/report.md`. Corpus replaced by
SAP-samples/abap-cheat-sheets (Apache-2.0): 370 docs, 2.81M real tokens. 45 docs (840k tokens, 30 %) are longer
than 16384 and removed. Train 309 docs / 1.89M tokens, valid 16 docs / 80k tokens (split by family, seed
20261003), 618 iterations for 2 epochs. `test.jsonl` = copy of valid (needed by `mlx_lm.lora --test`).
## Running (detached)
- MLX server PID 59352, log `runs/stage1/server.log`.
@@ -27,8 +32,7 @@ Task: `docs/stage1-training-task.md`. Settings and weights: `train/README.md`. U
1. B2: read the last lines of `runs/stage1/baseline.log`; summary from `runs/stage1/baseline.json`
(`t01_test_budget40` holds the T01 test result). Commit.
2. Step 1: `train/prepare.py` (real token counts with the base model tokenizer, length filter 16384,
95/5 split by document, `train/data/train.jsonl` and `valid.jsonl`, report). A4H is not needed.
2. Decisions for Kral on the data (see report): the 45 removed docs, DOC share, valid size.
3. Step 2, second part: valid loss of the base model (`mlx_lm.lora --test` without adapter; check the
options with `--help` first). Add it to `runs/stage1/baseline.json`.
4. Step 3 (training): Kral stops A4H; stop the MLX server; no other model loaded. Short test of 20

98
train/data/report.md Normal file
View File

@@ -0,0 +1,98 @@
# Stage 1 data report
Corpus: `/Users/erhankeseli/projects/abap-llm/corpus/corpus.jsonl` (source abap-cheat-sheets, Apache-2.0, see corpus/out/ATTRIBUTION.md).
Tokenizer: `/Users/erhankeseli/models/Qwen3.8-27B-4bit`. Seed 20261003. max_seq_length 16384.
Split by family (release versions and parts of one document stay together): 180 families.
| | docs | tokens |
|---|---|---|
| corpus | 370 | 2806501 (estimate chars/4: 2636112, ratio 1.065) |
| removed (> 16384) | 45 | 840136 |
| train | 309 | 1886475 |
| valid (= test file) | 16 | 79890 |
Iterations for 2 epochs (batch 1): 618.
## Token distribution per document (real tokenizer)
| set | min | p50 | p90 | p99 | max |
|---|---|---|---|---|---|
| all | 49 | 5195 | 16923 | 21483 | 31552 |
| kept | 49 | 4027 | 14998 | 16217 | 16273 |
| train | 49 | 4403 | 14998 | 16217 | 16273 |
| valid | 130 | 3392 | 16027 | 16129 | 16129 |
| tokens per document up to | docs |
|---|---|
| 512 | 48 |
| 1024 | 25 |
| 2048 | 40 |
| 4096 | 51 |
| 8192 | 53 |
| 16384 | 108 |
| more | 45 |
## By object type (kept documents)
| types | docs | tokens |
|---|---|---|
| DOC | 106 | 1012152 |
| CLAS | 147 | 827969 |
| PROG | 13 | 55365 |
| mixed(4) | 6 | 28405 |
| DDLS | 27 | 20614 |
| mixed(3) | 2 | 6667 |
| DDLX | 6 | 5003 |
| CLAS+DDLS | 1 | 4563 |
| FUGR | 2 | 2026 |
| INTF | 7 | 1792 |
| BDEF | 6 | 1564 |
| DCLS | 2 | 245 |
## Removed documents
- ZCL_DEMO_ABAP_OBJECTS (CLAS): 17042 tokens
- ZCL_DEMO_ABAP_STRUCTURES (CLAS): 16628 tokens
- 32_Performance_Notes__P03 (DOC): 16742 tokens
- ZCL_DEMO_ABAP_INTERNAL_TABLES (CLAS): 19117 tokens
- ZCL_DEMO_ABAP_INTERNAL_TABLES (CLAS): 18883 tokens
- ZCL_DEMO_ABAP_DYNAMIC_PROG__V755 (CLAS): 19801 tokens
- 23_Date_and_Time (DOC): 20347 tokens
- ZCL_DEMO_ABAP_INTERNAL_TABLES__V757 (CLAS): 17477 tokens
- 07_String_Processing__P02 (DOC): 17916 tokens
- ZCL_DEMO_ABAP_CONSTRUCTOR_EXPR (CLAS): 17368 tokens
- ZCL_DEMO_ABAP_DYNAMIC_PROG__V758 (CLAS): 31552 tokens
- ZCL_DEMO_ABAP_STRING_PROC__V757 (CLAS): 19499 tokens
- 04_ABAP_Object_Orientation__P06 (DOC): 16694 tokens
- ZCL_DEMO_ABAP_DTYPE_DOBJ__V758 (CLAS): 18368 tokens
- 06_Dynamic_Programming__P07 (DOC): 23994 tokens
- 28_Regular_Expressions (DOC): 17900 tokens
- 06_Dynamic_Programming__P06 (DOC): 16923 tokens
- ZCL_DEMO_ABAP_NUMERIC_OP__V816 (CLAS): 19769 tokens
- 05_Constructor_Expressions__P01 (DOC): 17349 tokens
- ZCL_DEMO_ABAP_DTYPE_DOBJ__V755 (CLAS): 17237 tokens
- ZCL_DEMO_ABAP_DTYPE_DOBJ__V756 (CLAS): 18069 tokens
- ZCL_DEMO_ABAP_DYNAMIC_PROG (CLAS): 17923 tokens
- ZCL_DEMO_ABAP_DYNAMIC_PROG (CLAS): 16924 tokens
- ZCL_DEMO_ABAP_DYNAMIC_PROG (CLAS): 16770 tokens
- 14_ABAP_Unit_Tests__P03 (DOC): 18549 tokens
- ZCL_DEMO_ABAP_STRING_PROC__V758 (CLAS): 22096 tokens
- 22_Released_ABAP_Classes__P01 (DOC): 19022 tokens
- 21_XML_JSON__P01 (DOC): 16625 tokens
- 29_Numeric_Operations__P01 (DOC): 17805 tokens
- 03_ABAP_SQL__P01 (DOC): 17305 tokens
- ZCL_DEMO_ABAP_DTYPE_DOBJ__V757 (CLAS): 18308 tokens
- 14_ABAP_Unit_Tests__P04 (DOC): 17633 tokens
- 24_Builtin_Functions (DOC): 16851 tokens
- ZCL_DEMO_ABAP_NUMERIC_OP (CLAS): 19740 tokens
- ZCL_DEMO_ABAP_INTERNAL_TABLES__V758 (CLAS): 17355 tokens
- ZCL_DEMO_ABAP_BUILTIN_FUNC (CLAS): 20446 tokens
- ZCL_DEMO_ABAP_STRING_PROC (CLAS): 20311 tokens
- ZCL_DEMO_ABAP_UNIT_TEST__V758 (CLAS): 16394 tokens
- ZCL_DEMO_ABAP_DYNAMIC_PROG__V757 (CLAS): 21483 tokens
- ZCL_DEMO_ABAP_DATE_TIME (CLAS): 18634 tokens
- ZCL_DEMO_ABAP_XML_JSON (CLAS): 16825 tokens
- ZCL_DEMO_ABAP_SQL (CLAS): 18283 tokens
- ZCL_DEMO_ABAP_INTERNAL_TABLES__V755 (CLAS): 17309 tokens
- ZCL_DEMO_ABAP_DYNAMIC_PROG__V756 (CLAS): 20979 tokens
- 16_Data_Types_and_Objects__P02 (DOC): 17891 tokens

370
train/data/stats.json Normal file
View File

@@ -0,0 +1,370 @@
{
"corpus": "/Users/erhankeseli/projects/abap-llm/corpus/corpus.jsonl",
"seed": 20261003,
"max_len": 16384,
"estimate_tokens": 2636112,
"real_tokens": 2806501,
"ratio_real_to_estimate": 1.065,
"all": {
"docs": 370,
"tokens": 2806501,
"min": 49,
"p50": 5195,
"p90": 16923,
"p99": 21483,
"max": 31552
},
"kept": {
"docs": 325,
"tokens": 1966365,
"min": 49,
"p50": 4027,
"p90": 14998,
"p99": 16217,
"max": 16273
},
"train": {
"docs": 309,
"tokens": 1886475,
"min": 49,
"p50": 4403,
"p90": 14998,
"p99": 16217,
"max": 16273
},
"valid": {
"docs": 16,
"tokens": 79890,
"min": 130,
"p50": 3392,
"p90": 16027,
"p99": 16129,
"max": 16129
},
"removed": [
{
"objects": [
"ZCL_DEMO_ABAP_OBJECTS (CLAS)"
],
"tokens": 17042
},
{
"objects": [
"ZCL_DEMO_ABAP_STRUCTURES (CLAS)"
],
"tokens": 16628
},
{
"objects": [
"32_Performance_Notes__P03 (DOC)"
],
"tokens": 16742
},
{
"objects": [
"ZCL_DEMO_ABAP_INTERNAL_TABLES (CLAS)"
],
"tokens": 19117
},
{
"objects": [
"ZCL_DEMO_ABAP_INTERNAL_TABLES (CLAS)"
],
"tokens": 18883
},
{
"objects": [
"ZCL_DEMO_ABAP_DYNAMIC_PROG__V755 (CLAS)"
],
"tokens": 19801
},
{
"objects": [
"23_Date_and_Time (DOC)"
],
"tokens": 20347
},
{
"objects": [
"ZCL_DEMO_ABAP_INTERNAL_TABLES__V757 (CLAS)"
],
"tokens": 17477
},
{
"objects": [
"07_String_Processing__P02 (DOC)"
],
"tokens": 17916
},
{
"objects": [
"ZCL_DEMO_ABAP_CONSTRUCTOR_EXPR (CLAS)"
],
"tokens": 17368
},
{
"objects": [
"ZCL_DEMO_ABAP_DYNAMIC_PROG__V758 (CLAS)"
],
"tokens": 31552
},
{
"objects": [
"ZCL_DEMO_ABAP_STRING_PROC__V757 (CLAS)"
],
"tokens": 19499
},
{
"objects": [
"04_ABAP_Object_Orientation__P06 (DOC)"
],
"tokens": 16694
},
{
"objects": [
"ZCL_DEMO_ABAP_DTYPE_DOBJ__V758 (CLAS)"
],
"tokens": 18368
},
{
"objects": [
"06_Dynamic_Programming__P07 (DOC)"
],
"tokens": 23994
},
{
"objects": [
"28_Regular_Expressions (DOC)"
],
"tokens": 17900
},
{
"objects": [
"06_Dynamic_Programming__P06 (DOC)"
],
"tokens": 16923
},
{
"objects": [
"ZCL_DEMO_ABAP_NUMERIC_OP__V816 (CLAS)"
],
"tokens": 19769
},
{
"objects": [
"05_Constructor_Expressions__P01 (DOC)"
],
"tokens": 17349
},
{
"objects": [
"ZCL_DEMO_ABAP_DTYPE_DOBJ__V755 (CLAS)"
],
"tokens": 17237
},
{
"objects": [
"ZCL_DEMO_ABAP_DTYPE_DOBJ__V756 (CLAS)"
],
"tokens": 18069
},
{
"objects": [
"ZCL_DEMO_ABAP_DYNAMIC_PROG (CLAS)"
],
"tokens": 17923
},
{
"objects": [
"ZCL_DEMO_ABAP_DYNAMIC_PROG (CLAS)"
],
"tokens": 16924
},
{
"objects": [
"ZCL_DEMO_ABAP_DYNAMIC_PROG (CLAS)"
],
"tokens": 16770
},
{
"objects": [
"14_ABAP_Unit_Tests__P03 (DOC)"
],
"tokens": 18549
},
{
"objects": [
"ZCL_DEMO_ABAP_STRING_PROC__V758 (CLAS)"
],
"tokens": 22096
},
{
"objects": [
"22_Released_ABAP_Classes__P01 (DOC)"
],
"tokens": 19022
},
{
"objects": [
"21_XML_JSON__P01 (DOC)"
],
"tokens": 16625
},
{
"objects": [
"29_Numeric_Operations__P01 (DOC)"
],
"tokens": 17805
},
{
"objects": [
"03_ABAP_SQL__P01 (DOC)"
],
"tokens": 17305
},
{
"objects": [
"ZCL_DEMO_ABAP_DTYPE_DOBJ__V757 (CLAS)"
],
"tokens": 18308
},
{
"objects": [
"14_ABAP_Unit_Tests__P04 (DOC)"
],
"tokens": 17633
},
{
"objects": [
"24_Builtin_Functions (DOC)"
],
"tokens": 16851
},
{
"objects": [
"ZCL_DEMO_ABAP_NUMERIC_OP (CLAS)"
],
"tokens": 19740
},
{
"objects": [
"ZCL_DEMO_ABAP_INTERNAL_TABLES__V758 (CLAS)"
],
"tokens": 17355
},
{
"objects": [
"ZCL_DEMO_ABAP_BUILTIN_FUNC (CLAS)"
],
"tokens": 20446
},
{
"objects": [
"ZCL_DEMO_ABAP_STRING_PROC (CLAS)"
],
"tokens": 20311
},
{
"objects": [
"ZCL_DEMO_ABAP_UNIT_TEST__V758 (CLAS)"
],
"tokens": 16394
},
{
"objects": [
"ZCL_DEMO_ABAP_DYNAMIC_PROG__V757 (CLAS)"
],
"tokens": 21483
},
{
"objects": [
"ZCL_DEMO_ABAP_DATE_TIME (CLAS)"
],
"tokens": 18634
},
{
"objects": [
"ZCL_DEMO_ABAP_XML_JSON (CLAS)"
],
"tokens": 16825
},
{
"objects": [
"ZCL_DEMO_ABAP_SQL (CLAS)"
],
"tokens": 18283
},
{
"objects": [
"ZCL_DEMO_ABAP_INTERNAL_TABLES__V755 (CLAS)"
],
"tokens": 17309
},
{
"objects": [
"ZCL_DEMO_ABAP_DYNAMIC_PROG__V756 (CLAS)"
],
"tokens": 20979
},
{
"objects": [
"16_Data_Types_and_Objects__P02 (DOC)"
],
"tokens": 17891
}
],
"by_type": {
"CLAS": {
"docs": 147,
"tokens": 827969
},
"mixed(4)": {
"docs": 6,
"tokens": 28405
},
"DOC": {
"docs": 106,
"tokens": 1012152
},
"INTF": {
"docs": 7,
"tokens": 1792
},
"BDEF": {
"docs": 6,
"tokens": 1564
},
"DDLS": {
"docs": 27,
"tokens": 20614
},
"mixed(3)": {
"docs": 2,
"tokens": 6667
},
"PROG": {
"docs": 13,
"tokens": 55365
},
"DDLX": {
"docs": 6,
"tokens": 5003
},
"CLAS+DDLS": {
"docs": 1,
"tokens": 4563
},
"DCLS": {
"docs": 2,
"tokens": 245
},
"FUGR": {
"docs": 2,
"tokens": 2026
}
},
"source": [
"abap-cheat-sheets"
],
"iterations_2_epochs": 618
}

113
train/prepare.py Normal file
View File

@@ -0,0 +1,113 @@
"""Stage 1 data preparation: real token counts, length filter, split by document, report.
train/.venv/bin/python train/prepare.py [--corpus PATH] [--max-len 16384] [--seed 20261003]
Output: train/data/train.jsonl, valid.jsonl, test.jsonl (copy of valid: `mlx_lm.lora --test` needs a test file),
train/data/report.md, train/data/stats.json. Only the tokenizer is loaded (no model).
"""
import argparse
import collections
import json
import os
import random
import re
from transformers import AutoTokenizer
ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
OUT = os.path.join(ROOT, "train", "data")
def pct(v, p):
return v[min(len(v) - 1, int(len(v) * p))]
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--corpus", default=os.path.expanduser("~/projects/abap-llm/corpus/corpus.jsonl"))
ap.add_argument("--model", default=os.path.expanduser("~/models/Qwen3.8-27B-4bit"))
ap.add_argument("--max-len", type=int, default=16384)
ap.add_argument("--seed", type=int, default=20261003)
ap.add_argument("--valid-frac", type=float, default=0.05)
a = ap.parse_args()
tok = AutoTokenizer.from_pretrained(a.model)
docs = [json.loads(l) for l in open(a.corpus)]
for i, d in enumerate(docs):
d["_id"] = i
d["real_tokens"] = len(tok(d["text"], add_special_tokens=False)["input_ids"])
est = sum(d["tokens"] for d in docs)
real = sum(d["real_tokens"] for d in docs)
kept = [d for d in docs if d["real_tokens"] <= a.max_len]
removed = [d for d in docs if d["real_tokens"] > a.max_len]
# Family = same demo in several release versions (__V755 ...) or parts of one text (__P01 ...). A family
# is never split between train and valid (near-duplicates would make the valid loss too good).
fam = collections.defaultdict(list)
for d in kept:
fam[re.sub(r"__(V\d+|P\d+)", "", d["objects"][0] if d["objects"] else str(d["_id"]))].append(d)
keys = sorted(fam)
random.Random(a.seed).shuffle(keys)
n_valid, valid, train = max(1, round(len(kept) * a.valid_frac)), [], []
for k in keys:
(valid if len(valid) < n_valid else train).extend(fam[k])
os.makedirs(OUT, exist_ok=True)
for name, part in (("train", train), ("valid", valid), ("test", valid)):
with open(os.path.join(OUT, name + ".jsonl"), "w") as f:
for d in part:
f.write(json.dumps({"text": d["text"]}, ensure_ascii=False) + "\n")
def summ(ds):
v = sorted(d["real_tokens"] for d in ds)
return {"docs": len(ds), "tokens": sum(v), "min": v[0], "p50": pct(v, .5), "p90": pct(v, .9),
"p99": pct(v, .99), "max": v[-1]} if v else {"docs": 0, "tokens": 0}
by_type = collections.defaultdict(lambda: [0, 0])
for d in kept:
k = "+".join(d["types"]) if len(d["types"]) <= 2 else "mixed(%d)" % len(d["types"])
by_type[k][0] += 1
by_type[k][1] += d["real_tokens"]
bins = [512, 1024, 2048, 4096, 8192, 16384, 10 ** 9]
hist = collections.Counter()
for d in docs:
for b in bins:
if d["real_tokens"] <= b:
hist[b] += 1
break
stats = {"corpus": a.corpus, "seed": a.seed, "max_len": a.max_len, "estimate_tokens": est,
"real_tokens": real, "ratio_real_to_estimate": round(real / est, 3), "all": summ(docs),
"kept": summ(kept), "train": summ(train), "valid": summ(valid),
"removed": [{"objects": d["objects"][:3], "tokens": d["real_tokens"]} for d in removed],
"by_type": {k: {"docs": v[0], "tokens": v[1]} for k, v in by_type.items()},
"source": sorted({d.get("source", "?") for d in docs}),
"iterations_2_epochs": len(train) * 2}
json.dump(stats, open(os.path.join(OUT, "stats.json"), "w"), indent=1)
L = ["# Stage 1 data report", "",
f"Corpus: `{a.corpus}` (source {', '.join(stats['source'])}, Apache-2.0, see corpus/out/ATTRIBUTION.md).",
f"Tokenizer: `{a.model}`. Seed {a.seed}. max_seq_length {a.max_len}.",
f"Split by family (release versions and parts of one document stay together): {len(fam)} families.", "",
"| | docs | tokens |", "|---|---|---|",
f"| corpus | {len(docs)} | {real} (estimate chars/4: {est}, ratio {stats['ratio_real_to_estimate']}) |",
f"| removed (> {a.max_len}) | {len(removed)} | {sum(d['real_tokens'] for d in removed)} |",
f"| train | {len(train)} | {stats['train']['tokens']} |",
f"| valid (= test file) | {len(valid)} | {stats['valid']['tokens']} |", "",
f"Iterations for 2 epochs (batch 1): {len(train) * 2}.", "",
"## Token distribution per document (real tokenizer)", "", "| set | min | p50 | p90 | p99 | max |", "|---|---|---|---|---|---|"]
for n in ("all", "kept", "train", "valid"):
s = stats[n]
L.append(f"| {n} | {s['min']} | {s['p50']} | {s['p90']} | {s['p99']} | {s['max']} |")
L += ["", "| tokens per document up to | docs |", "|---|---|"]
L += [f"| {b if b < 10 ** 9 else 'more'} | {hist[b]} |" for b in bins]
L += ["", "## By object type (kept documents)", "", "| types | docs | tokens |", "|---|---|---|"]
for k, v in sorted(by_type.items(), key=lambda kv: -kv[1][1]):
L.append(f"| {k} | {v[0]} | {v[1]} |")
L += ["", "## Removed documents", ""]
L += [f"- {', '.join(d['objects'][:2])}: {d['real_tokens']} tokens" for d in removed] or ["None."]
open(os.path.join(OUT, "report.md"), "w").write("\n".join(L) + "\n")
print("\n".join(L))
if __name__ == "__main__":
main()