Draft
Conversation
Integrate the Speaker (Sovereign AI Playbook) project as customization_recipes/ for language/domain model customization. What's added: - 6-stage customization pipeline (CPT, SFT, RL, BYOB, Eval, Quantization) - data_prep/ shared library (acquire, translate, SDG, quality, tokenize, BYOB, quantize) with lazy imports for optional deps - 12 SKILL.md agentic docs as primary agent interface - AGENTS.md repo capability map - 8 CLI commands under `nemotron customize` with shared _execute.py - Multi-container Docker deployment (orchestrator, curator, trainer, evaluator, NIM) with command dispatcher for single-entry-point UX - Airgap support (pre-download + deploy scripts + compose overlay) - Sovereign benchmark bridge (BYOB → NeMo Evaluator BYOB framework) - Lepton and Run:AI executor support in nemo_runspec/execution.py - Nemotron model family populated; Llama/Qwen stubs - [customize] optional dependency group (14 packages) - Tests for data_prep, configs, and SKILL.md integrity What's NOT changed: - Existing nano3/super3/embed recipes (zero modifications) - Core dependencies (customize deps are optional) - nemotron.kit, nemo_runspec (only execution.py extended) Verified: 8-point safety audit confirms no breakage to existing functionality. 3-round review cycle completed (all issues fixed). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove ~280 lines of reimplemented logic (tokenizer loading, file I/O, ShareGPT conversion, tokenization, packing, bin/idx building) and delegate to the existing production-grade nemotron.data_prep pipeline. tokenize_pack.py is now a thin adapter (272 lines, down from 549): - CPTConfig/SFTConfig dataclasses kept as OmegaConf interface - Added to_data_blend()/to_tokenizer_config() converters - prepare_cpt_data() delegates to run_pretrain_pipeline() - prepare_sft_data() delegates to run_sft_pipeline() - Thinking token support confirmed already in nemotron.data_prep (chat_template + nano3.jinja) — removed local reimplementation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Remove "(WIP)" from customize CLI entry in AGENTS.md — it is fully implemented with 8 commands - Add data quality evaluation (--mode data) documentation to the E2E SKILL.md stage 4 walkthrough Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Every SKILL.md now tells agents to gather requirements from the user BEFORE executing any commands. Adds structured input tables with specific questions to ask, defaults, and conditional triggers. - E2E SKILL.md: Step 0 with 7 required + 12 optional inputs - Per-stage SKILL.md: "Inputs Required" tables (7-12 inputs each) - data_prep SKILL.md: Per-utility input sections (acquire, translate, SDG, quality, tokenize/pack) Agents will no longer execute the Hindi medical example verbatim — they will ask the user what language, domain, data, and compute environment to use first. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Major changes: - Training CLI commands (cpt, sft, rl) now reuse existing nano3/super3 training scripts instead of maintaining copies. The --model-family flag (-m) selects which script set to use (default: nano3). - Added MODEL_FAMILY_SCRIPTS mapping and resolve_training_script() to _execute.py for dynamic script resolution at execution time. - Deleted duplicate run_cpt.py, run_sft.py, run_rl.py (replaced with tombstones pointing to the reused scripts). - YAML configs rewritten to Megatron-Bridge format (matching actual nano3/super3 configs): recipe._target_, train.*, model.*, etc. - All SKILL.md config references aligned with Megatron-Bridge keys. Additional fixes: - SDGConfig: added optional domain/language fields for targeted generation - acquire.py: auto-downloads FastText lid.176.bin when lid_model_path=None - deploy_airgap.sh: fixed service names to match docker-compose.yaml - CLI override examples: use correct Megatron-Bridge key paths - RL examples: use -c default with training_type= override - Env vars: standardized on OPENAI_API_KEY across docs and compose Usage: nemotron customize cpt -c default # nano3 (default) nemotron customize sft -c default -m super3 # super3 nemotron customize rl -c default training_type=dpo # DPO on nano3 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
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.
WIP code placement review. DO NOT MERGE.