Add Nemotron 3.5 embedding and reranking workflows (#351) * feat(retrieval): add multimodal Mistral3 recipes Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * fix(embed): record successful native mining provenance Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * fix(embed): include native metric logging dependency Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * refactor(embed): inherit checkpoint preprocessing across native stages Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * refactor(embed): use stage-local Git dependencies and extras Remove experimental wheel projects and project switching. Preserve text and VL dependency constraints as mutually exclusive stage extras. Use released DataDesigner and document the remaining bare-JSON handling limitation rather than carrying a core patch. Validation: 119 focused CPU tests; public Stage 0 Git dependency lock resolved. Remaining prep/train/eval locks and clean installation require publication of the local AutoModel source pin. No GPU or hosted generation rerun. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * fix(embed): lock public stage dependencies Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * Consolidate generic multimodal SDG and recipe integration Fold the grouped-query handoff, native encoder batch controls, BEIR compatibility fixes, and retrieval-first EA integration into PR #351. Pin DataDesignerPlugins #92 while preserving AutoModel #3915 and native mining behavior. Supersedes the separate #353, #356, #357, #358 and #359 stack. Signed-off-by: Steve Han * fix(embed): narrow multimodal EA and resolve integration regressions Signed-off-by: Steve Han * Expose generic multimodal SDG controls and pin completed producer Signed-off-by: Steve Han * docs: explain multimodal retrieval recipe input preparation Signed-off-by: Steve Han * Clarify multimodal EA configuration and publish its shared guide Signed-off-by: Steve Han * fix(embed): keep FlashAdamW optimizer states in FP32 Signed-off-by: Steve Han * fix(embed): retain FP32 parameters for unquantized optimizer states Signed-off-by: Steve Han * fix(embed): omit redundant FlashAdamW master-weight correction Signed-off-by: Steve Han * revert(embed): defer FP32 optimizer changes Restore the pre-FP32 FlashAdamW configuration, profiles, documentation and tests while a better precision change is designed and validated. Signed-off-by: Steve Han * Use generic semantic SDG planning in the multimodal EA recipe Signed-off-by: Steve Han * Pin the SDG producer with native abstention correction Signed-off-by: Steve Han * Lock summary clustering to the validated NumPy version Signed-off-by: Steve Han * Pin corrected bounded-context SDG reporting Signed-off-by: Steve Han * Default EA summary embeddings to the public Nemotron API Signed-off-by: Steve Han * fix(embed): make checkpoint deployment overrides opt-in Signed-off-by: Steve Han * docs(embed): record final-default end-to-end qualification Signed-off-by: Steve Han * fix(embed): pin optional quote diagnostics on development recipe Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * fix(embed): pin SDG source coverage and image grounding fixes Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * Pin whole-unit heading and TOC SDG section planner Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * Use rendered request limits and explicit multimodal model budgets Signed-off-by: Steve Han * Restore simple SDG defaults with optional model-token preflight Signed-off-by: Steve Han * docs(embed): link to the VLM SDG workflow Signed-off-by: Steve Han * fix(retrieval): update recipes to current Automodel main Pin embedding and reranking stages to 3914f200a4c782d44b58ee7a01b4685e4158e19c and regenerate their locks. Adapt native mining and text dataset interfaces, preserve checkpoint preprocessing, and retain adapter export dependencies. Validation: 521 recipe tests passed, 5 skipped; lint, formatting, and four lock checks passed. Two-GPU smoke covered native text/VL mining, two-step text/VL embedding and reranker training, reranker checkpoint resume, consolidated reload, and base/fine-tuned evaluation. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * feat(retrieval): add Nemotron 3.5 recipe workflows Add directory-based multimodal SDG, native embedding adaptation, and Nemotron 3.5 reranker fine-tuning. Update dependency pins, checkpoint retention, NIM truncation support, and public guides. Scope shared training adapters to retrieval recipes and make full export audits explicit. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * feat(retrieval): complete Nemotron 3.5 evaluation and serving Add native and vLLM reranker evaluation and saved-checkpoint deployment. Preserve multimodal processing, frozen retrieval candidates, and raw scores. Fix the model revision, writable serving caches, empty SDG views, and checkpoint timeout. Document JSON generation and checkpoint resources. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * fix(retrieval): handle prepared corpus paths and image-only evaluation Allow vLLM negative mining to reuse relative and absolute corpus references from prepared training files. Normalize missing BEIR text fields before dense retrieval over image-only documents. Cover re-mining and absent text fields, correct the evaluation handoff fixture, and select Docker GPUs explicitly in deployment examples. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * feat(retrieval): add repeatable validation and training resource checks Use one artifact root for Nemotron 3.5 rerank training, evaluation, and deployment while preserving explicit paths and existing text defaults. Report checkpoint disk and RAM estimates before model loading and reject clearly insufficient space for pinned model weights. Add a locked CPU recipe-contract CI job and a public mixed-modality verification runner for mining, training, resume, evaluation, and optional serving and 8B checkpoint checks. Document execution coverage and resource requirements without claiming a new full 8B validation run. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * fix(retrieval): validate checkpoint resources and training updates Recognize FlashAdamW's canonical class path after AutoModel serializes its configuration, so native training prints checkpoint resource estimates. Reject missing or nonfinite training metrics in the verification runner and retain the validated update in its checkpoint evidence. Clarify that GPU allocations and checkpoint staging share capacity on unified-memory systems. Validated with 368 CPU contracts, 19 native-runtime checks, strict docs builds, and the retained real embedding and 8B checkpoints from the Spark run. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * docs(skills): focus retrieval guidance and live evaluation coverage Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * Attach NVSkills validation signatures Signed-off-by: nvskills-svc-account * docs(skills): normalize retrieval metadata and credential guidance Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * fix(embed): preserve existing PyTorch index configuration Restore the Stage 0 cu129 source declaration and the embedding training CPU fallback for platforms other than Linux and macOS. These removals were unrelated to the multimodal integration. Regenerate the training lock without upgrading existing packages. Validation: both stage locks pass uv lock --check --offline. All eight original recipe PyTorch source mappings and index declarations match the PR base. The Stage 0 lock is unchanged; its transitive torch still resolves from PyPI. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * fix(sdg): default multimodal judging to Nemotron 3.5 Super VL Use nvidia/nemotron-3.5-super-vl-120b-a12b for summary and query judging in the Nemotron 3.5 profile, while retaining the judge environment override. Rename the generation variable to NEMOTRON_3_5_SDG_GENERATOR_MODEL and update the walkthrough, reference, skill guidance, and CLI field description. Validation: 92 focused tests pass, covering default and overridden judges through both configuration entrypoints. Sphinx builds with warnings as errors; Ruff and skill schema/quality checks pass. No hosted inference or live skill evaluation was run. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * test(embed): validate PyTorch sources by platform The restored training CPU fallback uses a list of sources. Replace assertions requiring a single source with marker evaluation for Linux, Windows, and macOS. Keep index URL and explicitness checks in the dedicated source test and remove the duplicate representation-dependent assertion from the stage-extras test. Validation: reproduced both GitHub failures locally, then all 370 CPU contract tests passed without skips. Ruff and git diff --check passed. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> * fix(sdg): separate multimodal model settings from text QA Add sdg_generator_model and sdg_judge_model for multimodal generation, with Super VL defaults resolved centrally before CLI preview and persistence. Preserve text QA model parameters and defaults. Keep the shared judge mapped to the plugin fallback so future per-stage overrides retain their native semantics. Validation: 386 CPU contract tests, 107 plugin integration tests, Ruff, strict Sphinx build, real CLI previews, and SkillEvaluator schema/static quality checks (100/100). No additional live evaluation cases. Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> --------- Signed-off-by: Oliver Holworthy <1216955+oliverholworthy@users.noreply.github.com> Signed-off-by: Steve Han Signed-off-by: nvskills-svc-account Co-authored-by: Steve Han Co-authored-by: nvskills-svc-account

Sunbelt Computer Software

PL/B Language Development and Support