erhan/abap-llm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
84 Commits 1 Branch 0 Tags
4432ac896c0ac7b0a669870c3a63aab3625fe9cd
Commit Graph

84 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
Kral
4432ac896c Only kinds below their target share are generated and run; one-controller flock guard 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 19:39:37 +02:00
Kral
976274d48a Balanced generation under the pipeline, per-kind brake; STATE: type mix fix 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 19:27:43 +02:00
Kral
55f4068330 STRU and exception tasks accepted; G6 cascade fix, exception class mutants, RTTI unit notes 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 19:25:09 +02:00
Kral
b119f1afac Object type mix: INTF, TABL, STRU, MSAG, exception tasks (harness G2, mutants, generator notes), balanced generator, kind-deficit job order, dashboard mix card 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 18:52:29 +02:00
Kral
454db6626c Review notes: token length for bf16 test (p95 48k), DDLS reject reasons, G1034 lock analysis (not reproduced) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 18:21:44 +02:00
Kral
224ba1a5f8 Summary: correct the guard line; pipeline summary reads the limit from .env 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 18:05:49 +02:00
Kral
5bcd1bf9cd Stage 2 summary at 50 accepted trajectories 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 18:02:52 +02:00
Kral
332fb4601d Live status page (SAP colors), rewritten every 5 minutes 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 17:56:30 +02:00
Kral
bc6827b744 Budget guard from panel 41.07: ratio 1.2, limit 121 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 17:47:39 +02:00
Kral
704cfa3ede Budget guard from the panel: ratio 1.35 for trajectory runs, limit 123 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 15:54:10 +02:00
Kral
ae37617a0d Budget guard: ratio 2.07 measured, limit 142; limit read from .env at each check 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 14:47:54 +02:00
Kral
c062df2ec5 Trajectory runs: tool-call budget 100 for tasks with a CDS contract object (eval keeps 60) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 14:08:33 +02:00
Kral
13360292dd Pipeline controller: plan 2 (hard +, error +50 %), backlog throttle, 2 trajectory workers, stop rules, summaries every 50; budget reserve 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 13:35:54 +02:00
Kral
1c4667a0fc MCP load test (read, heavy read, write modes) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 13:26:13 +02:00
Kral
9de3911578 K variants for training (free text, EPOD tool names), second attempt only for failed tasks, docs/epod-syntax-hint.md 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 12:50:43 +02:00
Kral
35f0eb0e0c Overlap limit for whole spec 0.75, slot logs; STATE and roadmap: stage 2 data progress 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 12:04:11 +02:00
Kral
61b903f7b3 Acceptance filter: score, end reason, harness errors, loop trimming, repair marker 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 12:03:54 +02:00
Kral
c2d4997566 Converter to the Qwen 3.8 chat template (thinking off) with tokenizer round-trip check 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 12:03:54 +02:00
Kral
d5e43e1a83 Trajectory record (messages, raw tool results, metadata; reasoning apart) and trajectory runner 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 12:03:54 +02:00
Kral
a4eb567e4c Proxy: local abaplint messages on a bare 'save failed' (EPOD syntax check cannot see the rejected source) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 12:03:54 +02:00
Kral
229f862600 Generator training mode: train pool, eval overlap check, category mix, error-targeted tasks 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 11:53:07 +02:00
Kral
3a9f9e75fc Step 400 check: Mac -13.4 % vs GPU -35.0 %; next GPU run bf16 mixed with stage 2; findings in STATE and roadmap 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-05 11:46:21 +02:00
Kral
39dcdd9846 Step 200 check: GPU -26.0 % vs Mac -10.7 %, full run cancelled at step ~418 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-04 20:48:49 +02:00
Kral
0c0982f666 Overfit conversion test passed (Mac -98.5 %, GPU -99.97 %); full run started, alpha 32 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-04 19:34:24 +02:00
Kral
20ef7b2331 Stage 1 pipeline test: 30.5 s/step, converter checked, valid loss 0.848 with test adapter 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-04 19:24:25 +02:00
Kral
5447874fd3 Stage 1 on HF Jobs: Unsloth job script, PEFT to MLX converter, base valid loss 0.849, Qwen base model docs 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
 2026-10-04 18:51:40 +02:00
Kral
a2ba9e7b44 serve.sh back to Qwen 3.8 27B 4-bit 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 17:02:59 +02:00
Kral
9f85a71ca5 Roadmap: Qwen baseline result 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 17:02:02 +02:00
Kral
a368233d26 Stage 1 reference baseline: Qwen 3.8 27B mean 15.8 (3/11 above 0), comparison with Devstral 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 17:01:52 +02:00
Kral
dbe6077a48 Decision 2026-10-04: base model Qwen 3.8 27B (Devstral not used further) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 13:11:53 +02:00
Kral
5532a5b226 baseline: setup failure is not recorded as a result; repair_stats tolerates a missing trajectory 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 12:07:33 +02:00
Kral
d15e190f72 baseline.py: --model, --base-url, --enable-thinking-false (Qwen run on MacBook) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 11:39:33 +02:00
Kral
0af2d6458e Devstral Small 2 baseline (11 tasks): README, roadmap, results; Qwen 3.8 dropped 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 11:20:43 +02:00
Kral
8d1db9c67c Devstral Small 2: serve.sh, baseline.py without thinking args, repair rate, activation error message fix, stop rule in chain (run base 21000) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 08:57:12 +02:00
Kral
e84ea43a3d Base model decision: drop Qwen 3.8, candidate Devstral Small 2 (24B) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 08:44:04 +02:00
Kral
ba0a7e6b3b STATE: correct start time 2026-10-04 07:59:41 +02:00
Kral
342eb6fc9a STATE: baseline restart time 2026-10-04 07:59:38 +02:00
Kral
5c68a1b1b7 Loop guard: also identical call with identical result 3 times in a row; baseline restarts (run base 20500) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 07:59:28 +02:00
Kral
a94c2a3b5d Baseline restarted with loop guard, thinking off (run base 20400) 2026-10-04 07:40:19 +02:00
Kral
0401186a6c Loop guard (3 identical pushes), end_reason and activation error records per run; thinking off for the stage 1 baseline 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-04 07:07:44 +02:00
Kral
040b9900fd STATE/handover: baseline restarted on the 11-task subset 2026-10-03 22:36:30 +02:00
Kral
689820af4d Baseline shortened: 11-task subset (1 per category + T01), max_tokens 16384, settings in README; baseline_chain.sh 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-03 22:36:00 +02:00
Kral
7054d9fd2e STATE: correct timestamps 2026-10-03 22:24:08 +02:00
Kral
7b8ca01bde Stage 1 data: strict version dedup (older v* vs next newer v*); counts updated 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-03 22:23:04 +02:00
Kral
0677d035da Stage 1 data: version dedup, splitting of long documents, family split; 748 iterations 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-03 22:09:15 +02:00
Kral
e083c9ca13 Stage 1 step 1: prepare.py, real corpus numbers (SAP-samples/abap-cheat-sheets) 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
 2026-10-03 21:58:04 +02:00
Kral
f0593933f9 Budget calendar: usage estimate, 20 USD paid for 60 USD usage 2026-10-03 20:15:05 +02:00
Kral
eaaef125ff Budget guard 135 (ledger), about 50 USD real 2026-10-03 20:12:38 +02:00
Kral
b0fd06259b Budget: real ratio ledger/2.7 (Ollama page 24.63 USD) 2026-10-03 20:11:45 +02:00
Kral
aef4c6e461 Pending: A4H memory limit after the baseline 2026-10-03 20:04:36 +02:00