rawmodels/unslop-harbor

Unslop (Harbor)

200 Harbor tasks with the reward we trained AlexWortega/Qwen3.8-27B-antislop on. Your model finishes a tweet or writes a HumanEval function body. Claude Opus 5.5 grades it against a rubric frozen for that task, ignoring length. A fixed rule handles length separately. Code has to pass its tests before the judge reads it.

Tasks	100 tweet continuations, 100 HumanEval bodies; half EN, half RU
Splits	8 train / 4 validation / 38 test per domain and language ([metadata].split)
Judge	anthropic/claude-opus-5.5 on OpenRouter, Anthropic as the only provider, reasoning low
Rubrics	3 or 4 criteria per task, written once by Opus 5.5 from the calibrated meta-prompts, stored in tests/task.json
Answer	the agent writes /workspace/answer.txt
Reward	a number in [0, 1] in /logs/verifier/reward.txt; criterion scores and the judge's evidence in grade.json

Reward

tweets:  r = L(words, target) · σ(a · (q − μ))
code:    r = q / 10 if every HumanEval test passes, else 0

• q averages the judge's 0–10 criterion scores. The judge runs the GEPA system prompt plus one clause telling it to ignore word count. We added the clause after a side-by-side audit: without it, verdicts followed length compliance in 18 of 23 decided tasks.
• L owns length, so a policy gains nothing by padding or trimming to please the judge. Inside ±15% of the target in the prompt, L = 1. Outside, it decays as exp(−(dev − 0.15)/0.35) and stays above zero, so in a group of overlong samples the shorter ones still win.
• μ and a turn q into the chance of beating a typical Qwen3.8-27B answer. We set μ to the judge's mean on 100 Qwen tweet continuations (6.74) and a to 1.7/σ (1.45).
• Code earns nothing until it passes, so the judge only ranks correct solutions. The grader runs the candidate as nobody with 3 s of CPU, 512 MB, 16 KB files, 64 descriptors and a 6 s wall clock.
• If the judge refuses, truncates or never answers, the grader writes no reward and Harbor logs an error instead of a
  0. An empty answer scores 0 without an API call.

Calibration

1. Judge prompt. GEPA searched the system prompt and the per-domain rubric meta-prompts on train tasks, treating "human beats model" as a prior. Validation picked the winner. Test tasks never reached the search.
2. Tweet rubrics. We tuned the tweet meta-prompt again on pairs with a known winner: the human tweet against the same tweet rewritten by Sonnet 5.5 as assistant prose. On 30 held-out tasks the judge picked the human text 98.3% of the time (seed prompt: 96.6%; 95% task-cluster CI [0.95, 1.00]). Code keeps the GEPA meta-prompt.
3. Reward. We fit μ and σ on judge scores of Qwen3.8-27B answers. tests/reward_config.json pins the judge prompt by judge_system_sha256, and tooling/build.py refuses to build if the hash drifts.

Layout

tasks/<domain>_<lang>_<id>/
├── task.toml           # judge env, timeouts, provenance (source, license, rubric model)
├── instruction.md      # the assignment word for word, plus where to write the answer
├── environment/        # Dockerfile: python:3.12-slim pinned by digest, unprivileged `agent` user
├── solution/solve.sh   # oracle: writes the human reference
└── tests/              # test.sh, grade.py, task.json (rubric, HumanEval tests), reward_config.json
data/tasks.jsonl        # one row per task: prompt, rubric, target_words, reference, source, license
data/reward_config.json # judge and calibration shared by every task
reward.py               # the same grader, importable outside Harbor
registry.json           # Harbor registry entry
manifest.json           # sha256 of every task directory
jobs/oracle.yaml, jobs/nop.yaml
tooling/                # build.py, gen_rubrics.py, rubric_metas.json

Run

hf download rawmodels/unslop-harbor --repo-type dataset --local-dir unslop
cd unslop
OPENROUTER_API_KEY=sk-or-... harbor run -p tasks -a <agent> -m <model> -e docker
# the validation runs below:
OPENROUTER_API_KEY=sk-or-... harbor run -y -c jobs/oracle.yaml
harbor run -y -c jobs/nop.yaml

For single-turn RL, skip Harbor. data/tasks.jsonl holds the exact prompts we calibrated on, without the /workspace/answer.txt line, and reward.py scores a completion:

import json, reward
cfg = json.load(open("data/reward_config.json"))
row = json.loads(open("data/tasks.jsonl").readline())
print(reward.grade({**row, "assignment": row["prompt"]}, cfg, completion_text)["reward"])

Validation

We ran every task through Harbor 0.15 on local Docker.

Agent	Trials	Errors	Mean reward
nop (writes nothing)	200	0	0.000 on every task
oracle (human reference)	200	0	0.776

Oracle subset	Mean reward	Mean q (0–10)
code EN	0.933	9.34
code RU	0.928	9.29
tweets EN	0.734	7.75
tweets RU	0.509	6.83

• All 100 HumanEval references pass inside the task image. Across 30 tasks, 69 of 85 single-site AST mutants fail; the other 16 compute the same thing as the reference (< → <= at a clamp boundary, for one).
• Every reference tweet hits the ±15% length band, so L = 1 and the preference term sets its reward.
• Human Russian tweets land near μ = 6.74. One calibration covers EN and RU tweets, and RuLexNorm tweets are noisy and informal.
• A judge call costs about $0.011: roughly 2k prompt tokens and 150 completion tokens.

Caveats

• We qualified the judge on catching degraded text in known-order pairs. Ranking two good answers is a harder test it hasn't passed yet.
• μ and a belong to Opus 5.5. Swap in another UNSLOP_JUDGE_MODEL and the grader still runs, with stale calibration.
• Calibration answers had thinking off. Opus 5.5 wrote these rubrics; GEPA ran with Muse Spark 1.1 as the rubric writer.
• The references come from public corpora: Broad Twitter Corpus, RuLexNorm, HumanEval. An agent with network access can find them. For RL, cut the agent's network or use the tasks as single-turn prompts.
• Sonnet 5.5 wrote the slop in the calibration pairs. Your policy's slop may look different.
• RU code tasks pair Russian instructions with the original English docstrings.
• We left out blogs and scientific abstracts, the other two unslop domains: their sources don't permit redistribution.

Credits

The method follows Meta FAIR's unslop / RL from eXpert Aligned Rubrics, rebuilt on a different model and judge. Harbor defines the task format and runs it. The layout copies the FineEnvs Harbor datasets.

License

Each subset keeps its source license ([metadata].license in every task, license in data/tasks.jsonl):

• EN tweets: GateNLP Broad Twitter Corpus, CC-BY 4.0 (README).
• RU tweets: Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm, CC-BY-SA 2.5 (README). This subset stays CC-BY-SA 2.5.
• Code: HumanEval, OpenAI, MIT (license).
• Rubrics, grader and task scaffolding: MIT.