Skip to content

docs: add MAI-Code-2-Flash fine-tuning support parity - #55

Open
John Wu (johnwu0604) wants to merge 1 commit into
mainfrom
johnwu0604-upgraded-broccoli
Open

John Wu (johnwu0604) wants to merge 1 commit into
mainfrom
johnwu0604-upgraded-broccoli

Conversation

@johnwu0604

Copy link
Copy Markdown
Collaborator

Summary

Add MAI-Code-2-Flash alongside MAI-Code-1.1-Flash with the same documented fine-tuning support. Reuse the existing code-repair recipe rather than introducing a second notebook.

  • Add the model to the root support table, managed catalog/demo metadata, model index, task index, learning path, and RFT track navigation.
  • Document the exact training ID mai-code-2-flash-2026-11-17 and catalog identifier Microsoft/mai-code-2-flash/2026-11-17, with an explicit notebook MODEL substitution. The date identifies a model version, not a launch date.
  • Apply the same managed-RFT-only methods, Global/Developer training types, Serverless Global Standard serving, grader prerequisites, and safety/evaluation guidance to both models. Continue linking to Microsoft Learn for region availability and REST/SDK/portal workflows.
  • Preserve the Code1.1 default, all executable notebook cells, datasets, grader definitions, and existing prices. No new prices, numerical limits, or quality claims are inferred from support parity.

Support parity baseline

Area Existing Code1.1 baseline Code2 documentation
Methods Root README.md; managed_training/MODEL_AND_NOTEBOOK_INDEX.md Managed RFT only; no SFT, DPO, or programmed training
Training and serving tiers Code-repair README.md setup/serving and notebook Sections 4/9 Global and Developer training; Serverless Global Standard serving; same unsupported tiers
Regions and API/SDK/UI workflows Existing links to the Learn support reference and RFT guide; notebook Sections 4–10 Same references and shared workflow; exact Code2 training ID supplied
Data, configuration and graders Code-repair metadata, preserved inputs and grader guide Shared schema, o3-mini-2025-01-31 grader and RFT configuration
Safety and result interpretation Code-repair result guidance, grader calibration warnings and notebook conclusion Same calibration, isolated-test/human-review and held-out-evaluation guidance

Validation

  • Root navigation/model-eligibility checks: 7 passed.
  • Repository workflow's selected managed offline contracts: 168 passed, 1 skipped. The optional installed-OpenAI-SDK serialization check is skipped because openai is absent from the minimal documentation-test environment; the mocked workflow runs for both model selections.
  • All 15 catalog/demo metadata entries synchronized; 42 relative links/anchors and 3 fenced Python examples checked.
  • Confirmed notebook executable cells, cell IDs, metadata and execution state are unchanged; final diff/whitespace review passed.

No live Foundry operations were performed. These checks validate documentation and local sample contracts, not service rollout/readiness, cloud availability, or held-out quality.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant