`reasoning_synth / scaler`¶

SCALER expands released parameterized problem families into concrete reasoning tasks. It uses the supplied generator and reference programs without an LLM call.

Pipeline, step by step¶

flowchart TD
  S["Family JSON: description, generator, references, scales"] --> D["Select family, difficulty and seeded sample"]
  D --> P["Scale parameters: int(scale * base + min)"]
  P --> G["Remote offline Python generator"]
  G --> I["Parse concrete input and problem detail"]
  I --> R["Run first successful Python / C++ reference"]
  R --> U["Reject duplicate concrete instance; record answer"]
  U --> B["Construct instruction and private answer-file verifier"]
  B --> H["Fresh Harbor: empty answer -1; reference +1"]
  H -->|"Pass"| O["Export algorithmic reasoning instance"]

Expand a family. Load native generator/reference definitions as data. Select a difficulty and random seed, map difficulty to a scale, and calculate each parameter. Supplied code is never executed on the controller.

Compute a concrete answer. Run the generator and reference in bounded offline remote containers. Record input, parameters, actual seed, code hash and answer. Duplicate concrete instances are rejected.

Create the learner contract. instruction_for combines the family description and concrete input. The learner writes a boxed answer to /workspace/answer.txt; a separate verifier applies the active native math-verify route.

Every prompt and its data¶

Zero LLM calls. All task content comes from released family definitions and deterministic program execution.

Call	System prompt composition	User / input material	Output	Retry or branch
No model prompt	instruction_for in scaler/families.py	Family description, concrete input dictionary and optional native instruction.	Learner instruction.md, assembled as text	This is task text, not an API call.

Read the complete scaler 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 hotel-room allocation family generates one sequence of arrival/departure operations. The reference computes that instance’s final allocation. The task asks the learner to compute the concrete answer; it is not a repository patch task.

What repeats, what is checked¶

The generator has at most max_generator_attempts (default three) per sampled instance. Supported references are Python and C++17. A reference timeout rejects the candidate. Harbor export requires native -1/+1 parity. This profile expands released families; it does not synthesize new families or reproduce adaptive training.

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-scaler.yaml. The input is a native mapping such as the released SCALER-data/train/SCALER-8.json, or the same format with user-supplied families. All supplied code executes inside offline containers on Modal or Daytona. The controller only reads JSON and packages artifacts.

difficulties, samples_per_difficulty, seed, target, max_candidates and execution bounds control the batch. Input hashes, actual random seeds, scaled parameters and reference-code hashes are recorded. Duplicate concrete instances are rejected. The released collection contains 100 distinct reasoning tasks from the recorded family bank; it does not claim 100 new families or repository coding tasks. See the release inventory.

The learner writes its boxed answer to /workspace/answer.txt. A separate verifier applies the active upstream math-verify metric and preserves its native −1/+1 reward. The reference answer never enters the learner image. Generation checks confirm execution and packaging; detailed quality acceptance comes later.

This profile expands existing families. Fresh-family synthesis, adaptive difficulty training and the verl training stack are outside its scope. See the shared CLI and cloud guide.

Credit: SCALER, Apache-2.0, commit 60c6c5037866c718f4c001ea338f9c5a91cb01ae. RFC 0026 and packaged recipes/scaler/provenance.md map the source and runtime adaptations.

reasoning_synth / scaler¶