bashgym

name	bashgym
description	Understand and operate the BashGym ML workspace across datasets, direct LLM training, terminal RL, evaluations, models, artifacts, campaigns, reports, and agent coordination. Use as the stable platform router; load the training or operator skill for executable work.
version	2.0.0
tags	["bashgym","training","evaluation","agents","artifacts"]

BashGym

BashGym is the shared ML workspace. It stores the authoritative operational record for datasets, model/config revisions, runs, metrics, evaluations, artifacts, budgets, and reports. Hermes, Codex, Claude Code, Discord, and canvas nodes are operator/interface choices over that workspace, not separate training systems.

The project-isolated experiment ledger extends the existing campaign SQLite database. Every official run carries workspace_id, project_id, experiment_id, run_id, attempt_id, exact model/dataset/environment version IDs, evaluation/artifact IDs, source identity, and correlation ID. Query one project at a time; never infer a project from the most recent conversation.

Route the task

• For session planning, authority, monitoring, reporting, resumption, and GBrain curation, read ../bashgym-operator/SKILL.md.
• For selecting, launching, monitoring, stopping, or evaluating a training method, read ../training/SKILL.md.
• Before a direct LLM launch, read ../training/references/bashgym-launch-recipes.md.
• Before activating local, SSH, or cloud compute, read ../training/references/compute-target-activation.md.
• For the general architecture boundary, read references/architecture-overview.md.
• For evidence and promotion gates, read references/eval-capabilities.md.
• For a multi-iteration baseline/hypothesis loop, use the durable campaign API, bashgym campaign doctor, and campaign ledger surfaces; do not start new research on the prototype /api/autoresearch/* surface.

Do not treat this router as live run state. Inspect the API/CLI, run manifests, runtime processes, and current GBrain sources before stating what is active or what happened most recently.

Direct training boundary

POST /api/training/start, bashgym training start, and the start_training agent tool directly support:

• SFT;
• DPO;
• GRPO;
• RLVR;
• teacher distillation;
• Session Distillation.

DPPO replay/backends, ECHO/RWML diagnostics, reward-model training, and cascade/MOPD use separate method-specific workflows. Embedding retriever training is one project/profile lane, not the default product architecture and not a causal-LM training alias.

compute_target must select a real execution lane, not just label provenance. Direct BashGym runs accept local, ssh:<device_id>, and cloud:nemo-customizer. NeMo RL runs through registered private compute. Hugging Face Jobs, managed provider fine-tunes, and SkyPilot/dstack plans use separate surfaces and retain their upstream job identity.

Artifact defaults

Routine direct runs should begin with:

{
  "checkpoint_limit": 1,
  "artifact_retention": "adapter_only",
  "auto_push_hf": false,
  "hf_private": true,
  "hf_upload_artifact": "auto"
}

Choose adapter_checkpoint only for completed runs that must remain resumable, deployable only for an intentional standalone serving artifact, and full_run only for audited/promoted work. Public Hugging Face upload requires explicit release authority.

Non-negotiable evidence

• Save the exact strategy/config and dataset/model lineage before launch.
• Use list_experiment_projects, get_experiment_context, and get_experiment_run (or bashgym ledger ...) to recover current project state.
• Pass a complete training tracking context for official work. Missing context is recorded under unassigned, not attached to a guessed project.
• A smoke run proves the runtime path, not model quality.
• A loss curve is training evidence, not a promotion decision.
• Use heldout, environment, reward/replay, safety, baseline-comparison, and RunCard evidence appropriate to the method.
• Keep raw logs, high-volume metric series, datasets, checkpoints, and model files in BashGym artifact storage. Curate concise decisions, milestones, findings, and artifact references into GBrain.
• Consume ledger events by durable cursor when curating GBrain or an optional cloud sink. Do not make a product Supabase database a required BashGym backend.
• Do not merge product repositories to share a model. Exchange versioned artifacts and explicit integration contracts.