DR-OPIC

DR-OPIC is a runnable Python framework for coding SLM experiments:

Domain-Routed On-Policy Iterative Correction.

It does three concrete things:

1. Runs student coding attempts against executable tests.
2. Builds verified repair/delta/preference training records from real failures.
3. Reports the metrics that matter for a coding SLM: greedy@1, coverage@K, selected@K, selector_gap, and repair@1.

No private datasets, PDFs, model weights, or Kaggle outputs are included in this repo.

Install

From the repo root:

python -m pip install -e ".[dev]"

Run the test suite:

python -m pytest -q

Expected result:

7 passed

Run The Built-In Demo

The demo uses one toy Python task. The first student answer fails, the second passes, a repair candidate is verified, and DR-OPIC emits rollout, repair, ZPD, advantage, winner, and delta-span records.

python -m dr_opic.cli forge-demo

You can also use the installed console command:

dr-opic forge-demo

Write a stable artifact bundle:

python -m dr_opic.cli --output outputs\demo forge-demo

This creates:

round_summary.json
student_rollouts.jsonl
verified_repairs.jsonl
learnable_winner.json
delta_spans.json

Verify A Python Candidate

Create a task JSON:

{
  "prompt": "Implement reverse_words(s) returning words in reverse order.",
  "entrypoint": "reverse_words",
  "tests": "assert reverse_words('one two three') == 'three two one'\nassert reverse_words('solo') == 'solo'",
  "task_id": "reverse_words_demo"
}

Run verification against code embedded in the JSON or from a file:

python -m dr_opic.cli verify-python examples/python_task.json --code examples/reverse_words_good.py

Expected output includes:

{
  "passed": true,
  "observation": "passed"
}

Compute ZPD Weight

python -m dr_opic.cli zpd --passes 2 --samples 5

This prints Jeffreys-smoothed pass rate and the ZPD weight:

p_tilde = (passes + 0.5) / (samples + 1)
w_zpd = 4 * p_tilde * (1 - p_tilde)

Route And Estimate

Route a prompt through the domain and safety checks:

python -m dr_opic.cli route "Fix this Python traceback and add pytest coverage"

Estimate dense model memory and per-token compute:

python -m dr_opic.cli estimate-model --params 3.09e9

Build a counterfactual delta-span record:

python -m dr_opic.cli delta --task-id reverse_words --failed examples\reverse_words_bad.py --fixed examples\reverse_words_good.py

Run the verifier-ZPD scheduler demo:

python -m dr_opic.cli schedule-demo

Audit A JSONL Training File

The repo does not ship training data. To audit an SFT JSONL stored outside this repo:

python -m dr_opic.cli audit-jsonl C:\datasets\slm\sft.jsonl --schema sft

Expected SFT schema:

{"prompt": "...", "response": "..."}

For preference rows:

python -m dr_opic.cli audit-jsonl C:\datasets\slm\preferences.jsonl --schema preference

Required preference fields:

{"prompt": "...", "chosen": "...", "rejected": "..."}

Modules

• dr_opic.maths: ZPD, rewards, advantages, coverage metrics, cost estimates.
• dr_opic.verifier: Python code extraction, static checks, test execution.
• dr_opic.forge: student-first rollout, repair, and artifact construction.
• dr_opic.selectors: verified learnable-winner selection.
• dr_opic.delta: token/line delta spans between failed and fixed code.
• dr_opic.scheduler: verifier-ZPD curriculum buckets and train-mix weights.
• dr_opic.preference: scalar helpers for verified DPO/ORPO-style pairs.
• dr_opic.datasets: JSONL schema and quality audit helpers.
• dr_opic.replay: deterministic replay certification.
• dr_opic.routing: domain routing and abstention helper.
• dr_opic.safety: simple coding-safety acceptance helper.
• dr_opic.compression: memory/compute estimates and retention gates.
• dr_opic.losses: optional PyTorch losses for SFT, delta, DPO, and RLVR.

More detail:

• Quickstart
• Architecture
• API Reference
• Math Notes
• Release Protocol
• Notebook Tutorial

Safety Scope

verify-python executes candidate code in a temporary subprocess with a timeout. That is useful for local research, but it is not a security sandbox. Run untrusted model code inside a container or VM.