`terminal_synth / dataarc`¶

This DataArc-inspired recipe augments complete Harbor tasks with few-shot, self-instruct, in-depth evolution and in-breadth evolution. Recipe version 2 keeps those stages and input options, using Repo2RLEnv-authored prompt wording.

The September expansion completed 100 Harbor tasks (20 retained, 80 new), published as HuggingEnvs/repo2rlenv-dataarc. All 80 new exports have hash-matched baseline reward 0 and reference reward 1. The expansion accounted for \(26.57**, approximately **\)0.33 per new export, including unsuccessful model attempts and estimated worker/build costs. The published manifest separates those controls from independent quality acceptance, which remains unestablished. These results describe version 1. Version 2 has contract tests, but has not yet been evaluated in a paid generation campaign; the figures above are not a measurement of the replacement prompts.

Pipeline, step by step¶

flowchart TD
  S["Existing Harbor seed tasks"] --> L["Bounded seed excerpts + full environment context"]
  L --> ST["Enumerate strategy, direction and sample index"]
  ST --> P1["P1 · Complete artifact variant"]
  P1 --> D["Validate TerminalDraft"]
  D --> Q["Q1 · Optional draft consistency review"]
  Q --> E["Emit temporary Harbor bundle"]
  E --> R["Remote fresh nop + oracle"]
  Q -->|"Blocking issue within max_repairs"| P1
  D -->|"Schema feedback"| P1
  R -->|"Bounded execution feedback"| P1
  R -->|"0 / 1"| H["Export child with strategy and parent lineage"]

P1, P2, … identify actual model calls. Unlabelled stages are code or remote execution.

Read a seed. Selected instruction, reference and test excerpts provide the task context; environment files are also supplied in full within the supported text-size bound. Native canary-line filtering is retained and documented in provenance.

Enumerate transformations. few_shot makes a close variant; self_instruct makes a related task; evol_instruct chooses in_depth or in_breadth. Samples are independent variants of the selected parent, not a chained curriculum.

Generate complete artifacts. The deterministic design function wraps the seed. The first and only authoring stage creates a complete TerminalDraft. Retries use the same strategy and seed with real execution feedback.

Every prompt and its data¶

The artifact author can be followed by Q1 consistency review when review_drafts: true. This adds a model call on each complete draft before execution. Its blocking findings return to the same bounded artifact-author loop.

One artifact-author call on the first attempt. There is no separate design LLM call.

Call	System prompt composition	User / input material	Output	Retry or branch
P1 · Artifact / repair	artifact_prompt.md with strategy text from strategies.json + common materialization + owned adaptation	Selected seed excerpts are substituted into the system template; user JSON contains environment_context and feedback.	TerminalDraft	Initial call plus max_repairs; default three attempts.

Read the complete dataarc prompt reference for every retained template, appended instruction, substitution, example and output schema. The shared prompt guide explains how to inspect the fully resolved request from a real run.

Follow one task¶

Illustration: a scheduling seed becomes a related scheduling problem with a new constraint. The generated environment, solution and tests must agree on that constraint and preserve the seed’s actual tools.

What repeats, what is checked¶

Retries rebuild the same variant from the original seed context plus previous failure evidence. Null evolution direction for non-evolution strategies is omitted from TOML metadata. Full independent validation is deferred; strategy labels are not quality scores.

An exported bundle is a generation result. Independent leakage review, shortcut probes and blind solver traces belong to the later quality campaign.

Implementation map¶

Run and supported profile¶

Run repo2rlenv generate --config examples/owned-dataarc.yaml. The input directory contains seed task subdirectories, each with task.toml, instruction.md and solution/solve.sh. The first profile accepts text assets in a single CPU Linux container. Native examples cover portfolio optimization, asynchronous task cancellation and constraint scheduling. Users may supply their own Harbor seeds.

strategies, evol_directions and samples_per_strategy control enumeration. Each variant is authored directly from its seed. There is no separate design model. The seed's domain and tools must survive any execution-informed repairs. Parent hashes and strategy names remain in the generated lineage.

Shared options control target, max_candidates, max_repairs, token limits and test timeouts. The initial 20-task target has expanded to 100 generated tasks. Baseline failure and reference success are generation checks; detailed quality validation follows the full generation campaign. See the remote execution and CLI guide.

Credit: DataArc-SynData-Toolkit, terminal branch 2a1d65ec8dcfaea2458d67e1fb18078cce6420b9. That revision has no recorded license grant. Version 2 replaces the previously retained prompts and removes the license from a different branch. The method remains credited; existing version 1 artifacts retain their historical provenance. See RFC 0022 and provenance.md for the source and licensing boundary.

terminal_synth / dataarc¶