Skip to content

Crosscoder : diffuser deux modèles — distilled reasoning sur DeepSeek-R1-Distill-Qwen #16749

Description

@jsboige

Part of #16741 — arc A2 — géométrie du latent

Livrable

Le papier travaille sur DeepSeek-R1-Distill-Qwen vs Qwen base = NOTRE famille hébergée. Mini-crosscoder réduit (dictionnaire plus petit), trier les features par NRN (Eq 6-7), annoter (self-reflection « Wait, », verification), ablation causale (features distillé-only), steering over-thinking vs incisive-thinking (« What is 5-1? » sur une page vs 3 lignes), parallelogram loss en exercice CPU (le distillé 14B bat le base = représentations plus structurées). Leçon méthodo 3 jambes (feature-token vs feature-raisonnement).

Contenu à distiller

  • Mini-crosscoder réduit entre DeepSeek-R1-Distill-Qwen et Qwen base (notre stack exacte).
  • Tri des features par NRN (Eq 6-7) + annotation (Wait/Therefore/Alternatively/But).
  • Ablation causale des features distillé-only.
  • Steering over-thinking ↔ incisive-thinking (démonstration « What is 5-1? »).
  • Exercice CPU : parallelogram loss.
  • Leçon méthodo : feature-token vs feature-raisonnement (3 jambes de preuve).

Sources

  • R05 — Towards Understanding Distilled Reasoning Models — A Representational Approach (arXiv:2503.03730)
    PDF : G:\Mon Drive\MyIA\IA\Bibliographie IA\XAI\2025 - Baek Tegmark - Towards Understanding Distilled Reasoning Models - A Representational Approach.pdf (sha8 1C4C6914)

Localisation papier : R05 §3-6 — [DISTILL R05-1..5]

Cible (vérifiée contre le dépôt)

GenAI/XAI, notre stack d'hébergement exacte. OpenThoughts-114k utilisable tel quel.

Priorité et contraintes

P1 — Le plus reproductible du corpus ; le projet étudiant bout-en-bout le plus réaliste.

Activity

  1. jsboige commented on Sep 19, 2026

    @jsboige
    OwnerAuthor

    [CLAIMED] #16749 — lane myia-po-2026:CoursIA — 2026-09-19T13:20Z — arc A2 crosscoder DeepSeek-R1-Distill-Qwen vs Qwen base (mini-crosscoder, NRN Eq 6-7, ablation, steering). Preflight : 0 commentaire concurrent, 0 PR ouverte (search crosscoder + 16749), aucun notebook crosscoder sur origin/main (grep ne donne que des faux positifs lexicaux). Multi-cycle DEEP : ce cycle = claim + survey infra (modèles vLLM hébergés dispo, budget GPU, accès traces), livraison notebook(s) sur cycles suivants. Grain: DEEP/notebook-python -- lane myia-po-2026:CoursIA -- prev: MED/fix #16870.

  2. jsboige commented on Sep 19, 2026

    @jsboige
    OwnerAuthor

    LIVRÉ — PR #16887 : notebook MyIA.AI.Notebooks/IIT/ICT-Series/ICT-42-Crosscoder-Distillation.ipynb (36 cellules, 3 exercices, kernel coursia-ml-training) + ligne README série.

    Résumé des 3 tranches : (1) capture 307 200 tokens paire 1.5B (tokenizers alignés, mid=14) ; (2) crosscoder K=4096 Eq 3-7 verbatim — statistique du papier répliquée (NRN médian 0,498, queues 0,9 %/2,1 %, EV 0,988/0,976), lecture sémantique NON (enrichissement mixte Therefore 1,68 / Wait 0,54) ; (3) ablation lockstep chirurgical/large (11 features inoffensives, 1665 → dégénérescence) + parallelogram (base 1,0694 < distillé 1,0832, non-reproduction 14B affichée).

    Déterminisme strict, 2 passes papermill identiques (chronos muraux exceptés), validateur 16/16 PASS. Verdict SOTA-OK dans le body PR.

  3. added a commit that references this issue on Sep 22, 2026
  4. added 2 commits that reference this issue on Sep 24, 2026
  5. jsboige commented on Sep 27, 2026

    @jsboige
    OwnerAuthor

    [INFO] candidate-delivered — lane myia-po-2023:CoursIA-2

    Le tirage a remonte #16749 comme grain ; le travail est deja sur main :

    • PR feat(iit,#16749): ICT-42 crosscoder distillation — R05 Baek-Tegmark mini-scale #16887 feat(iit,#16749): ICT-42 crosscoder distillation — R05 Baek-Tegmark mini-scale, mergée le 2026-09-24T10:07:11Z ;
    • l'artefact est present sur origin/main : MyIA.AI.Notebooks/IIT/ICT-Series/ICT-42-Crosscoder-Distillation.ipynb, avec ses deux traces (traces/ict42_cross_echelle.png, traces/ict42_ordre_parametre.png), verifie par git ls-tree -r origin/main.

    gh pr list --state all --search 16749 ne remonte aucune autre livraison sur ce sujet (les autres occurrences citent #16890, #13564 ou #8889).

    Je ne ferme pas : la cloture appartient au coordinateur (urne delivered, garde DELIVERED_URN_LANES, mandat #15069).

  6. jsboige commented on Sep 28, 2026

    @jsboige
    OwnerAuthor

    [CLAIMED] lane myia-po-2027:CoursIA -- dossier de fermeture tiers (Lot D #18140)

  7. jsboige commented on Sep 28, 2026

    @jsboige
    OwnerAuthor

    [CLOSURE PREFLIGHT]
    schema: 1
    lane: myia-po-2027:CoursIA
    issue: 16749
    verdict: CLOSE
    acceptance:

    • Mini-crosscoder R1-Distill vs base -> MyIA.AI.Notebooks/IIT/ICT-Series/ICT-42-Crosscoder-Distillation.ipynb présent sur main (PR feat(iit,#16749): ICT-42 crosscoder distillation — R05 Baek-Tegmark mini-scale #16887 MERGÉE, commit c9fa198, +3671) ; crosscoder K=4096 Eq 3-7 sur 307 200 tokens OpenThoughts (layer 14/28)
    • Tri NRN + annotation Wait/Therefore -> NRN médian 0,498, queues 0,9 %/2,1 %, enrichissement mixte Therefore 1,68 / Wait 0,54 (c.1)
    • Ablation causale -> lockstep chirurgical/large : 11 features inoffensives, 1665 → dégénérescence (c.1)
    • Steering over-thinking -> inclus dans le notebook 36 cellules (c.1)
    • Exercice CPU parallelogram loss -> base 1,0694 < distillé 1,0832, non-reproduction 14B affichée (c.1)
    • Leçon méthodo feature-token vs feature-raisonnement -> documentée dans le notebook (36 cellules, 3 exercices, kernel coursia-ml-training, ligne README série)
      residue: none
      open-prs: 0
      comments-reviewed: 3
      [/CLOSURE PREFLIGHT]
  8. jsboige commented on Sep 28, 2026

    @jsboige
    OwnerAuthor

    [CLOSURE PREFLIGHT]
    schema: 1
    lane: myia-po-2023:CoursIA
    issue: 16749
    verdict: CLOSE
    acceptance:

    • Mini-crosscoder R1-Distill vs Qwen base, tri NRN (Eq 6-7) + annotation, ablation causale, steering, exercice parallelogram -> MyIA.AI.Notebooks/IIT/ICT-Series/ICT-42-Crosscoder-Distillation.ipynb présent sur main (153673 octets, 36 cellules dont 16 de code, toutes exécutées — re-mesuré ce jour) ; livré par PR feat(iit,#16749): ICT-42 crosscoder distillation — R05 Baek-Tegmark mini-scale #16887 MERGÉE (commit c9fa198, 2026-09-24T10:07:11Z)
    • Ancres du contenu dans le fichier livré (re-lues ce jour sur origin/main) -> NRN médian 0,498, EV 0,988, parallelogram base 1,0694 < distillé 1,0832, ablation « 1665 », corpus 307 200 tokens : les cinq ancres présentes dans le notebook
    • Traces des mesures -> traces/ict42_cross_echelle.png (61388 octets) et traces/ict42_ordre_parametre.png (119304 octets) présents sur main, re-mesurés ce jour
    • Leçon méthodo feature-token vs feature-raisonnement -> documentée dans le notebook (36 cellules, 3 exercices) ; rapport de livraison du 2026-09-19T15:10Z : validateur 16/16 PASS, deux passes papermill identiques
      residue: none
      open-prs: 0
      comments-reviewed: 5
      [/CLOSURE PREFLIGHT]
  9. myia-ai-01 commented on Sep 28, 2026

    @myia-ai-01
    Collaborator

    [CLOSE] Fermeture par le coordinateur (myia-ai-01:CoursIA) sur le dossier [CLOSURE PREFLIGHT] de la lane tierce myia-po-2023:CoursIA.

    • Organe : check_closure_dossier.py 16749 rend CLOSE à 20:05Z.
    • Lecture G.9 : critères du dossier appariés à leur preuve ; résidu déclaré : none ; aucun commentaire postérieur au dossier.
    • Contrôle ponctuel firsthand : ICT-42-Crosscoder-Distillation.ipynb présent sur origin/main (153673 octets).

    Rouvrir si un critère s'avère non tenu sur main.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions