feat(post-training,#12653): PT-11c GRPO-RLVR cran 1.7B/2B structure notebook - #12885
Merged
Merged
Conversation
…otebook Grain CONTENU frais c.556 -> c.557 : creation structurelle de PT-11c sur base PT-11b, avec adaptations Qwen3-1.7B/2B (cran au-dessus 0.8B) : - Architecture compatible trl.GRPOTrainer + peft 0.20.0 (post-install c.556) - 4 seeds x 100 steps (vs 2x20 PT-11b) -> intra-seed DM informatif >=22 pts - Verdict memoire cible RTX 3080 Ti 16GB (pic VRAM mesure) - Verifier SymPy Tier-1 + Z3 N-queens Tier-2 repris verbatim PT-11b - rewardspy.watch_trl wrappe en fallback no-op si absent (env po-2026) - LOAD_MODEL_AND_TRAIN=False par defaut (mode lecture, code C.1-compliant) - 3 exercices C.1 : parser etendu sci, dataset geometrie, cross-cran 11b/11c - 33 cellules (18 md + 15 code), execution end-to-end 0 erreur Acceptance #12653 : - (1) structure notebook cree : OK ce cycle - (2) run live 4x100 steps : c.558 - (3) verdict honnete BEATS/NO BEATS/MECANISME_REPRO/INCONCLUSIVE + PR : c.559 README PostTraining MAJ : pedagogical_count 14->15, breakdown PostTraining=15, maturite BETA=10 + ALPHA=5, ligne PT-11c inseree entre PT-11b multi-seed et PT-12. Notebook commite avec outputs reels (C.2) - C.1 sans erreur volontaire. Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com>
jsboige
commented
Aug 25, 2026
jsboige
left a comment
jsboige
left a comment
Owner
Author
There was a problem hiding this comment.
[Hermes] — review notebook-tier (structure PT-11c), vérifications firsthand sur le diff.
- Authenticité des outputs (règle C.2) : 15/15 cellules code exécutées,
execution_count1→15 contigu, 0 null. Les streams sont un vrai env probe : Python 3.11.9/Windows, torch 2.13.0+cpu,CUDA dispo : False, peft 0.20.0, trl 1.10.0, z3 4.16.0, sympy 1.14.0, etrewardspy : NOT INSTALLED (no-op wrapper...)— chaque claim d'env du body a son pendant littéral dans les outputs. Structure-only assumé : aucun résultat de training fabriqué (run live = c.558, hors scope déclaré). - Règle C.1 : 0
NotImplementedError/assert False/1/0; les 3 exos étudiants sont des stubsreturn None+TODO etudiantconformes. - Comptes : 33 cellules (18 md + 15 code) confirmées ; README
pedagogical_count: 14→15,PostTraining=15,BETA=10, ALPHA=5— arithmétique cohérente avec l'insertion d'une ligne. - Security scan : 1 match =
tokenizer.pad_token = tokenizer.eos_token— faux positif (assignation pad token, pas un credential). 0 vrai match (HF_TOKEN|API_KEY|BEARER|PASSWORD|SECRET|TOKEN=).
Structure conforme à la spec #12653, acceptances 2-3 explicitement gated sur c.558/c.559 — posture honnête. (Contrainte token : COMMENT only, auteur = jsboige.)
Contributor
Contributor
MD hierarchy drift -- 6a62105Cette PR augmente le compte de defauts de rendu markdown Corriger (ex. |
Contributor
Golden-Set Execution (H.7 P3)✅ 8/8 notebooks passed (certified reproducible)
Pinned lockfile: |
Owner
Author
|
[CONTEXTE REVIEW c.563] PR #12885 — MD hierarchy drift advisory : faux positif scanner (c.549-L1 ★ NEW pattern) ContexteLane Diagnostic : FAUX POSITIF SCANNER (c.549-L1 ★ NEW)La PR #12885 ne touche AUCUN des 3 notebooks signalés. Vérification firsthand Seuls fichiers modifiés :
Aucun des 3 notebooks signalés ( Tell = c.549-L1 ★ NEWLe scanner MD hierarchy drift ne filtre pas le diff par PR : il compare la baseline (burndown) au corpus complet, et tout delta — même sur des fichiers que la PR ne touche pas — est attribué à la PR. Pattern mesuré c.549 sur #12825 ( Note :
Note : le commentaire Action
Refs
Lane : myia-po-2026:CoursIA-2 |
Contributor
MD hierarchy drift -- 5353f8cCette PR augmente le compte de defauts de rendu markdown Corriger (ex. |
Contributor
MD hierarchy drift -- 0bffacfCette PR augmente le compte de defauts de rendu markdown Corriger (ex. |
Contributor
Notebook PR Validation: PASS
Checks: H.1 (no errors), H.3 (execution_count), C.1 (no banned patterns) |
This was referenced Aug 26, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
feat(post-training,#12653): PT-11c GRPO-RLVR cran 1.7B/2B — structure notebook
Grain: DEEP/training -- lane myia-po-2026:CoursIA-2 -- prev: LIGHT/refactor (cycles narrow sustained c.512+)
Résumé
Création structurelle du notebook
PT_11c_grpo_qwen17_rlvr.ipynbsur base PT-11b, avec adaptations Qwen3-1.7B/2B (cran au-dessus 0.8B). Cycle c.557 : structure ; c.558 : run live 4 seeds × 100 steps sur RTX 3080 Ti 16GB ; c.559 : verdict honnête + post-merge livraison.Livrables
MyIA.AI.Notebooks/GenAI/PostTraining/PT_11c_grpo_qwen17_rlvr.ipynb(33 cellules, 18 md + 15 code)Architecture
trl.GRPOTrainerextract_answer_sympy+math_verifier_rewardrewardspy.watch_trl(optionnel) +rlvr_reward_func#10603loss_fn='linear'informatif_torch.cuda.max_memory_allocated(0) / 1e9Acceptances #12653
linear+ delta > 0Verdict 4 classes (cf
pr-review-discipline.md §C) :linearp<0.05 ET delta > 0Règles respectées
raise NotImplementedError/assert False/1/0, stubspass/return Noneconformes (3 exos)Hors-scope ce cycle
Voir aussi
sensitivity_lean+knot_leannon-buildable — pas concerné ici (PT-11c est training Python)See #12653 (livraison partielle — issue reste ouverte pour c.558/c.559)