Own-Data Scenarios

Versioned source shapes that Halo Forge can inspect, map, validate, and hand to a compatible trainer

Halo Forge’s guided own-data workflow is driven by a versioned scenario registry. A scenario describes the minimum source fields, safe aliases and constants, canonical Dataset Lab shape, compatible trainer modes, and a small launch budget or proof recommendation where applicable. It does not bypass data validation, backend preflight, or the explicit launch step.

Choose Train on your data in the desktop or browser dashboard, or use the same Dataset Lab catalog from the CLI. See Use Your Own Data for the complete source-to-version workflow.

Available Scenarios

The stable registry anchor is shown exactly as returned by Halo Forge. All currently published scenarios are revision 1.

Scenario ID	Registry anchor	Canonical shape	Registry trainer modes
`instruction-sft`	`own-data/instruction-sft`	SFT	SFT
`chat-sft`	`own-data/chat-sft`	Chat	SFT
`preference-pairs`	`own-data/preference-pairs`	Preference	DPO, ORPO, RM
`prompt-reward`	`own-data/prompt-reward`	Prompt	RAFT, GRPO
`reasoning-sft`	`own-data/reasoning-sft`	SFT	SFT
`tool-agentic`	`own-data/tool-agentic`	Tool	SFT, agentic
`vlm-captioning`	`own-data/vlm-captioning`	VLM	VLM
`vlm-qa`	`own-data/vlm-qa`	VLM	VLM
`audio-asr`	`own-data/audio-asr`	Audio	audio
`corpus-adaptation`	`own-data/corpus-adaptation`	Corpus	CPT
`text-classification`	`own-data/text-classification`	Classification	classify
`text-multilabel`	`own-data/text-multilabel`	Classification	classify
`embedding-pairs`	`own-data/embedding-pairs`	Embedding	embed
`reranking`	`own-data/reranking`	Reranking	rerank
`image-classification`	`own-data/image-classification`	Classification	classify
`audio-classification`	`own-data/audio-classification`	Classification	classify
`audio-tts`	`own-data/audio-tts`	Audio	unavailable

audio-tts is intentionally returned as unavailable. Halo Forge has no verified data-to-weight-update contract for that task. Specialized task-model scenarios use the verified PyTorch runtime and are filtered out when their model, dependency, or backend contract is unavailable.

Use Help me decide or halo-forge data scenarios advise to receive ranked, explainable suggestions from a plain-language goal and the observed source shape. Halo Forge still requires the operator to confirm the scenario.

Registry-Matched Examples

These JSONL records match examples in the checked-in scenario registry and are parsed by the release-interface check.

Instruction SFT (instruction-sft@1):

{"instruction":"Summarize the water cycle.","answer":"Water evaporates, condenses, and returns as precipitation."}
{"instruction":"Name the capital of Japan.","answer":"Tokyo."}

Preference pairs (preference-pairs@1):

{"prompt":"Explain gravity simply.","chosen":"Gravity pulls masses toward one another.","rejected":"Gravity is when things always fall down."}

Prompt reward (prompt-reward@1):

{"problem":"What is 12 * 8?","reference_answer":"96"}
{"problem":"Return the first prime after 10.","reference_answer":"11"}

Tool and agentic (tool-agentic@1):

{"messages":[{"role":"user","content":"Weather in Austin?"},{"role":"assistant","content":"","tool_calls":[{"name":"weather","arguments":{"city":"Austin"}}]}],"tools":[{"name":"weather","parameters":{"type":"object","properties":{"city":{"type":"string"}}}}]}

VLM caption manifest (vlm-captioning@1):

{"image":"images/sample.png","caption":"A red bicycle beside a brick wall."}

Audio transcription manifest (audio-asr@1):

{"audio":"clips/hello.wav","transcript":"Hello from Halo Forge."}

Corpus adaptation (corpus-adaptation@1) uses a real document fixture rather than a labeled JSONL record:

# Dataset versions

Completed versions are immutable. Refreshing a changed source creates a new
source revision.

The inline rows require the referenced assets. Halo Forge’s Try a working example action and halo-forge data scenarios template ... --output DIRECTORY include tiny checksummed PNG/WAV companions, so the downloadable fixtures are train-path-ready. Replace them with real local assets for useful training; Halo Forge resolves, hashes, decodes, and validates those assets before building a version.

Small Managed Launch

For a local file, inspect first and then use the same scenario-backed defaults:

halo-forge data inspect --path ./support-sft.jsonl --scenario instruction-sft
halo-forge data add --name support-sft --path ./support-sft.jsonl \
  --scenario instruction-sft \
  --map prompt=instruction --map response=answer \
  --accept-recommended
halo-forge data show <dataset-id>
halo-forge data build <dataset-id> --recommended-recipe
halo-forge data render <version-id> --trainer sft
halo-forge sft train --dataset-version <version-id>

Labeled scenarios can recommend a bounded proof budget, but the final command still runs the selected trainer. Corpus adaptation uses an explicit CPT token/pass budget instead. Verify the semantic preview, quarantined rows, split policy, hardware preflight, and output path before launching.

Prompt-only RAFT/GRPO proof runs additionally require a compatible qualified verifier revision. The dashboard offers qualified profiles by name; raw verifier configuration remains an Advanced, unqualified launch.

For corpus adaptation:

halo-forge data extract --path ./manuals
halo-forge data inspect --path ./manuals --scenario corpus-adaptation
halo-forge data render <version-id> --trainer cpt --model Qwen/Qwen2.5-1.5B
halo-forge cpt train --dataset-version <version-id> \
  --model Qwen/Qwen2.5-1.5B --adaptation lora \
  --budget-mode passes --corpus-passes 1

Execution Surfaces

Desktop, local browser, remote browser, and CLI use the same source catalog, scenario revisions, immutable Dataset Lab versions, and work queue. A remote browser does not make its own filesystem visible to the Halo Forge workstation; upload a supported file or enter a path that exists on the workstation. See Workstation Surfaces for the platform and distribution matrix.