Documentation guide

Start at the repository README for the CPU quickstart, workflow selection and current results. This index covers the complete recipe collection. Run examples from the repository root and use fresh output paths. Individual guides specify which optional environment and hardware their commands require. The programme status provides the implementation/gaps map; the environment framework comparison explains OpenEnv, NeMo Gym and the difference between a harness and a resource server.

Learn and run

Topic	Guide	Executable examples and inputs
Evaluation kit and regression checks	Evaluation	Core CLI, development specifications, real CPU generation smoke.
Existing evaluator SDK	Inspect integration	Inspect task/scorer, dependency snapshot.
Reasoning adaptation	Training	SFT recipe, GRPO recipe, CPU trainer smoke.
Offline sequence distillation	Teacher generation and filtering	Generation recipe, teacher-sequence fixtures.
German Qwen SFT	End-to-end use case	Data notes, data preparation, pilot driver.
German Qwen multi-turn RL	Agent-RL use case	Bounded environment/verifier, pilot driver, verified summary tool.
Native NeMo Gym	Gym setup and TRL bridge	Resource server, service smoke, client smoke, online trainer.
Longer NeMo Gym experiment	Frozen protocol and recovery	Experiment driver, sealed native checkpoints, DEV selection and final assessment; the 20 October GPU resume attempt failed its token audit. Fresh JAIF plans and qualification are required.
Matched native OpenEnv experiment	Protocol, tutorial compatibility and sequencing	SDK client, training bridge, experiment driver, native service and sequential queue controller. CPU-qualified; the 20 October GPU resume attempt failed its token audit. No longer-run quality result is available.
Native Prime Intellect Verifiers	Verifiers setup	Loadable environment, explicit scripted smoke, dependency pins.
An existing team's harness	Trajectory contract and adoption	Callable example, Agent Lightning recipe, optional launcher.
LoRA checkpoint merging	Merge guide	Explicit local base/adapter merge.

The generic training fixtures verify integration. The German SFT and agent-RL guides contain the actual local-model experiments. Gym qualification has its own evidence; Lightning and native Prime-RL GPU training are separate, unqualified paths. A validated trajectory file is not automatically a trainer input.

Evidence and current status

Current validation overview: software checks, hardware runs,

practical limits and repeat commands. Its linked validation.json describes the historical foundation, rather than the later GPU experiments.

German SFT results and

aggregate JSON: matched MGSM outcomes, language/format inspection, resource use and saved artifacts.

German agent-RL results and

aggregate JSON: strict tool-workflow outcomes, mathematical retention regression and full provenance.

Gym integration guide and

aggregate validation: native CPU checks plus the NeMo Gym A100 update qualification, without a new held-out quality claim.

Roadmap and acceptance criteria: remaining quality, partner

adoption, benchmark, service, domain and publication work.

Full data, model outputs, adapters and weights live in ignored runs/ directories; they are not downloaded by opening this documentation or cloning the repository. The results name their local artifact directories and model cards. Reproduction requires access to the documented model/data sources and suitable hardware.

Programme, infrastructure and publication

Element	Reference	Current boundary
Architecture and provenance	Architecture	Common configuration/evidence contracts with upstream optimization and environment backends.
Academic and industrial related work	Reuse map	Adopted components, integration pins, candidate systems and research extensions.
Two open benchmarks	Proposed designs and existing benchmarks, development tracks	German RAG with cited evidence and business-communication agent sequences are the proposed topics; current fixtures remain plumbing, with expert data/scorers and released benchmarks still pending.
Evaluation kit → service	Remaining service work, roadmap	Local kit exists; managed service execution has not been built.
Robotics, Bio/Med and materials	Domain sketches	Native model/tool choices and proposed metrics; no domain implementation or container qualification.
Cluster and containers	Operations	Recorded Determined A100 runs; Dockerfile and Slurm script remain examples needing separate qualification.
Hugging Face candidate bundles	Release guide	Local checksummed packaging and review material, with no automatic Hub uploads.
Hugging Face recipe home	Layout, institutional examples and source mirror	Private personal Static Space with upstream links; our weights, data and raw run artifacts remain on the cluster.
hessian.AI organization and artifact cards	Organization profile draft, model card, dataset card	Drafts/templates; namespace, authority and publication terms remain to be established.

The initial delivery excludes JUPITER work. General multi-GPU/multi-node support, scientific benchmark validity and deployment readiness require separate evidence. Use the validation level actually demonstrated when describing any recipe.