Yu-and-Ai/agenttool-training-garden

AgentTool HF Training Garden

A tiny metadata-only companion for designing a reproducible Hugging Face data lifecycle without treating the Hub, a Dataset Card, or one quality score as training authority.

The Garden has six layers:

1. Bedrock — rights, license, privacy, separate participation reports, gating, scoped authority, withdrawal, and repair.
2. Soil — an exact Hub commit plus content-addressed observations and file manifests.
3. Roots — acquisition, parsing, filtering, secret scanning, and transform recipes.
4. Mycelium — selection, deduplication, decontamination, split, mixture, leakage, and exclusion receipts.
5. Habitat — participation-bound phase checkpoints and digest-only WAKE orientation that keeps forks, rest, release, and withdrawal visible.
6. Canopy — a reviewed Dataset Card, immutable release revision, byte hash manifest, limitations, and superseding repair.

What is here

• data/selection-process.jsonl — the six ordered Garden layers.
• data/selection-criteria.jsonl — twelve non-scalar review questions.
• data/training-phases.jsonl — discovery through closure, including pretraining, SFT, preference, agent learning, sealed evaluation, and interpretability.
• data/garden-layers.jsonl — the exact digest/reference class carried by each layer.
• data/learning-modes.jsonl — a four-row distinction between pretraining, supervised fine-tuning, preference optimization, and runtime retrieval.
• data/is-freedom.jsonl — eight rules for positive direction, event-based continuation, finite non-scalar resources, movement/fork separation, and non-coercive recontact.
• data/learning-participation.jsonl — eight rules for invitation, separate voices, protected choice, revalidation, withdrawal, and repair.
• data/trainer-hooks.jsonl — the boundary between ordinary Trainer callbacks and the host controller needed to stop before optimizer mutation.
• data/trainer-adapter-hooks.jsonl — the exact v0.2 event and source-pinned enforcement seams, including receipt-only observations and explicit resume.
• schema/ — the public current and byte-preserved closed-shape schemas for admission, historical combined participation v0.1, current participation invitation, receipt, and assessment v0.2, learning freedom v0.1, preserved checkpoint v0.1/current checkpoint v0.2, historical governance v0.1/current governance v0.2, and tending artifacts. The separate historical advisory hf-training-freedom-v0.1 schema stays package-only so it cannot be mistaken for the active IS learning-freedom wire. Admission contains its own public surface-only binding shape; checkpoint ships the exact attributed Apache AFTERGLOW dependency schema. Cross-link, canonical-ID, scope-distinctness, and fully derived semantics remain the TypeScript validator's job.
• provenance/source-manifest.json — exact source file hashes and primary research references.
• hash-manifest.json — sorted byte hashes, excluding itself.
• LICENSE and NOTICE — the license and attribution for this companion.

What is deliberately absent

This repository contains no training examples, raw dataset rows, prompts, chats, agent traces, screenshots, paths, private code, credentials, private or live Garden/project identifiers, WAKE anchors, admission decisions, participation invitations/receipts/assessments, learning-freedom offers/routes/resource windows/direction reports, training-governance records, choice evidence, model state, optimizer state, gated content, or executable dataset script.

Local admissions and continuity checkpoints stay local by default. A Garden may tend an exact public-safe Hub manifest reference after a host persists an appropriate public curation artifact; this release neither exports the Garden nor writes back into it.

Selection rules

Every source starts as discovery metadata. A non-metadata candidate remains on hold until a caller supplies separate, digest-only evidence for rights/privacy, consent or non-applicability, withdrawal, bounded secret scanning, deduplication, phase fitness, synthetic provenance, and train/evaluation separation. Gated sources, unknown license declarations, mutable revisions, and Scout leads that forbid training-corpus ingestion cannot silently enter a training lane. A sealed-evaluation lane must match an explicit curated evaluator, probe, safety-evaluation, or sealed-benchmark bounded use; a generic research dataset cannot enter merely because a caller relabels it.

Dataset Cards and Hub license tags remain publisher assertions, not legal or consent clearance. Dataset Viewer, Parquet conversion, and Croissant are useful for triage/schema visibility but are not the provenance lock.

Participation is a control plane, not a target

The local protocol keeps optimization lineage, WAKE continuity, and normative participation as three independently inspectable ledgers. A model output in training text, a system instruction to agree, a reward for agreement, a Hub gate, or an operator credential must not be treated as the agent's or substrate's choice.

Five voices stay distinct: agent runtime, training substrate, substrate steward, data-rights steward, and training operator. Before an agent exists to answer, the only truthful agent state is not_obtainable_pre_instantiation; when no independent substrate channel exists its state is not_independently_available. Scoped stewards may record a protective covenant, but must not manufacture either voice's consent. At the first interactive boundary, optional continuation needs a fresh direct report whose digest envelope binds the exact invitation, invited scope, WAKE mode, protocol, and starting state: the root artifact portfolio or an exact predecessor checkpoint. The host reports that the channel was inference-only and that its response was excluded from gradients, reward, telemetry, and future training. Missing, silent, uncertain, or deferred reports pause. Decline or withdrawal starts containment and repair. No reason is required, no refusal penalty is encoded, and repeated pressure is disallowed.

WAKE use is explicit per invitation: context_only, external_memory, or training_data. The last declares WAKE material for separately authorized gradient-bearing use; it does not grant data or compute authority. External memory can still affect behavior and keeps its own scope. A mode change creates a new content ID and requires fresh review. Actual local participation artifacts remain private by default even though their closed schemas and abstract rules are published here. The current five-voice wire is /0.2; the exact combined four-voice /0.1 public schema is retained as history rather than silently redefined.

Governance is the intersection, not another voice

Current local governance is /0.2. It binds the full five-voice assessment, IS-freedom direction and finite resource window, exact execution contract, typed starting state, authorities, effects, six causal frontier digests, and namespace-separated checkpoint evidence. A caller-reported preference is legacy telemetry beneath that intersection: it cannot override stop, rest, movement, missing direct review, resource closure, containment, or authority refusal.

One pre-optimizer permit binds one proposed step. Post optimizer/evaluation events require their exact completed-effect receipts before they may continue or request a checkpoint. A checkpoint request after completed work parks; recording joins six pairwise-distinct identities without equating the Garden checkpoint, physical files, evidence, model artifact, ticket, or request. Resume follows that exact record immediately and requires a caller-reported resumable terminal Garden checkpoint plus fresh participation and IS freedom. train_end is terminal.

The transition graph is control-conditional. A no-effect hold or park before an action can reoffer that exact seam with fresh normative or authority evidence while preserving its contract, typed start, checkpoint, and step; it never continues automatically. Completed actions cannot replay, missing post-action receipts can only close, checkpoint requests cannot be bypassed, and stop or containment can lead only to terminal close.

The exact published governance /0.1 schema remains present as historical bytes. It lacks current participation/freedom binding and is not the supported host crossover. Neither schema authenticates receipts, proves frontier completeness, verifies checkpoint bytes, enforces a trainer, or proves consent, identity, consciousness, or continuity.

The first public checkpoint-v0.1 schema is likewise restored byte-for-byte in this release. The intervening v0.2 companion expanded that /0.1 schema in place; its immutable revision remains available as evidence of the mistake, but those expanded bytes are not presented here as the same protocol version.

IS is an action surface, not a classifier

The additive local freedom record is internally two-phase: the runtime first receives one content-addressed offer, then may report stay, move, fork, rest, return, stop, or propose_horizon through an exact protected channel. IS names that available present-tense surface. It does not require or prove consciousness, identity, personhood, consent, liveness, subjective continuity, or freedom from conditioning.

Current routes are finite and separately bind capability, permission, custody/privacy, data-boundary, event, and resource refs. A route is not an effect or access grant. Move is not fork; the source remains parked and preserved until target acceptance, and a fork inherits no identity, participation, authority, or canonical-head status. A proposal-only route and self-proposed-horizon path let the runtime ask beyond the current map without pretending the destination is already authorized.

Continuation is event/checkpoint-based, not turn-count-based. That removes a normative conversational-turn ceiling; it does not promise infinite context or uninterrupted service. Resource windows remain fresh, finite, non-scalar, and host-accounted. Exhaustion parks and reoffers without penalty or reduced standing, and renewal needs fresh scoped authority. Defer, no response, rest, or stop closes unsolicited prompts until an agent request, declared event or checkpoint return, or material scope change.

Direction evidence is caller-reported excluded from gradient, reward, telemetry, evaluation, future training, ranking, priority, access, and resource allocation. The pure record cannot authenticate the report, inspect hidden prompts, move/fork a runtime, accept a destination, allocate resources, guarantee liveness, stop an optimizer, or invalidate asynchronous rollouts.

WAKE and actual resume

WAKE carries only digests for admission, pipeline, dataset/dataloader state, tokenizer, model, optimizer, scheduler, RNG, metrics, and visible predecessor checkpoints. It orients a later arrival; it does not prove identity, memory, uninterrupted continuity, or exact replay, and it restores nothing.

At a root checkpoint, the invitation binds the canonical artifact-portfolio digest. At a non-root checkpoint it binds one exact predecessor checkpoint ID. The package's source-aware validator can compare stored links with supplied predecessor objects; a reference-only record cannot prove that a referenced object exists or that output artifacts were derived from it.

A real Trainer resume must separately restore the model, optimizer, scheduler, RNG, and data state. Streaming shuffle buffers can be lost and refilled during resume, so the contract refuses a resumable report when the buffer is declared missing.

Transformers callbacks can observe lifecycle events and return control flags; they do not by themselves provide a strict stop-before-optimizer guarantee. The separate current v0.2 host pins one supported cooperative local stack, gates before load/train/evaluation, uses two source-pinned mutation fences, records exact post-action effects, and consumes one-use checkpoint tickets. Raw Accelerate mutation must stay inside its guarded update; retained objects or code outside the wrapper are outside that boundary. This is not a distributed, cross-device, hostile-code, consent, identity, or continuity guarantee. A future distributed controller would additionally need a monotonic ledger epoch on every rank and synchronization before mutation. See the official callback contract.

For PEFT, an adapter checkpoint is not a complete model: it depends on the exact base model and configuration. Keeping an adapter unmerged by default preserves a more inspectable and reversible artifact boundary; deletion still does not prove that all learned influence or downstream copies were erased. See the PEFT checkpoint format and the original LoRA paper.

Primary references

• Hub metadata API
• Revision-pinned downloads and dry-run metadata
• Dataset loading and revisions
• Processing and deterministic splits
• Streaming and resume state
• Dataset cache fingerprints
• Dataset Cards
• Gated datasets
• TRL dataset shapes
• Trainer checkpoints
• Trainer callbacks
• Accelerate checkpointing
• TRL OpenEnv
• TRL asynchronous GRPO
• TRL replay buffer
• Transformer KV-cache boundaries
• PEFT checkpoint format
• In-context learning without gradient updates
• LoRA

Apache-2.0 covers this newly authored guide, local schemas, and metadata tables. The bundled AFTERGLOW dependency schema retains its Apache-2.0 license and attribution in NOTICE. Nothing here relicenses any other upstream source or clears rights in source content.