rldoctor

Your GRPO run isn't broken. It's dead, and still logging.

Point rldoctor at a training log and it tells you which failure you have, how much it has cost you so far, and what to change.

The problem

RLVR runs fail silently. The loss curve stays smooth, reward keeps ticking up, and the job runs to completion — while the thing you actually wanted stopped happening a thousand steps ago.

Three examples, all of which your dashboard already has the data to detect and none of which it will tell you about:

• Your groups are degenerate. In GRPO, when every rollout in a group earns the same reward, the centred advantage is exactly zero for all of them. You paid for G generations and bought no gradient. TRL logs this as frac_reward_zero_std and shows it to you as a line on a chart. At 75% it means you are paying for four rollouts per useful rollout.
• Exploration is gone. Policy entropy decays in every run, so a low value proves nothing on its own. What matters is arrival time: entropy heading for the floor inside your remaining step budget, while reward has already plateaued.
• Your verifier is being gamed. An audit of code RL environments found that 28.5% of SWE-bench Verified tasks have test suites weak enough to accept a Docker-verified incorrect patch (arXiv:2606.16062). Against a grader that permeable, "reward went up" is not evidence of anything.

Every one of these is well documented in the literature and reproduced in a dozen papers. None of them is packaged as something you can run against your own job.

Try it in ten seconds

No GPUs, no account, no training run:

pip install rldoctor
rldoctor demo reward_hacking

Or the latest from main: pip install git+https://github.com/junglezke/rldoctor.

   CRIT  Reward-eval divergence (verifier gaming)
        Training reward is climbing while the held-out score is flat. The policy is
        optimising something your verifier rewards and your task does not.

        evidence
          · normalised training reward has moved +2.63 from its start; held-out
            score has moved +0.13 -- hacking gap +2.25
          · rank correlation between training reward and held-out score: rho=-0.05
            over 17 evaluations
          · the gap is widening (+9.61e-03/step, p=0.000463)

        do this
          1. Read 20 high-reward rollouts end to end. Reward hacks are almost always
             obvious on inspection and almost never visible in aggregate metrics.
          2. Audit the verifier, not the model. For code tasks, check whether your
             tests accept known-wrong patches; for rubric rewards, check for keyword
             stuffing.
          3. Hold out a second verifier that grades the same task differently
             (isomorphic verification). A gap that exists against one grader and not
             another localises the exploit to the grader.
          4. Add an explicit anti-hack penalty for the specific exploit once you
             have identified it, and re-measure the gap rather than assuming it
             closed.
          5. Roll back to the checkpoint before the gap opened -- later checkpoints
             have already been shaped by the exploit.

The banner above shows a different run: one where 75% of the rollout budget produced zero gradient. advantage_collapse does not just report that number — it separates all-correct from all-wrong. Same symptom, opposite fix, and the bare fraction cannot tell you which one you have.

Twelve scenarios ship with the tool (rldoctor demo --list), each reproducing a documented failure mode.

Point it at a real run

# Nothing to instrument — read a run you already logged
rldoctor diagnose wandb://my-team/my-project/3xk91abc

# Or a HuggingFace checkpoint, verl's file logger, a TensorBoard directory, a JSONL, a CSV
rldoctor diagnose ./checkpoint-1200/trainer_state.json
rldoctor diagnose ./my_project/my_experiment.jsonl     # verl: trainer.logger='["console","file"]'
rldoctor diagnose ./outputs/grpo-qwen7b/tensorboard
rldoctor diagnose ./logs/train.jsonl --format html -o report.html

From Python:

from rldoctor import diagnose, load_run

d = diagnose(load_run("wandb://my-team/my-project/3xk91abc"))
print(d.headline)
# [CRIT] Reward-eval divergence: Training reward is going up while the held-out
# score goes *down*. This is the textbook signature of reward hacking. (+2 more)

for f in d.problems:
    print(f.severity.label, f.title, "->", f.prescription[0])

Catch it during the run, not in the post-mortem

A post-hoc report tells you that you wasted 40 GPU-hours. A live monitor stops you.

from rldoctor.live import LiveMonitor

monitor = LiveMonitor(config=cfg, check_every=25)
for step, metrics in training_loop():
    monitor.log(metrics, step=step)     # logs a warning the first time something breaks

On TRL, that is one line:

from rldoctor.integrations.trl import RLDoctorCallback

trainer = GRPOTrainer(..., callbacks=[RLDoctorCallback(report_path="report.html")])

On the bundled entropy_collapse scenario, the live monitor flags the run at step 124 of 400 — with a projected step at which exploration runs out, and the DAPO clip-higher fix, before three quarters of the budget is spent.

To be precise about what that is: it is during, not before. The live monitor sees a failure once it is statistically visible in the metrics, which is usually well before anyone reads the dashboard and well after the onset. Whether scalar training metrics can give genuine early warning of reward hacking at a usable false-positive rate is an open question — independent work proposing onset hooks for verl reports that these signals reliably audit hacking after the fact but do not reliably warn ahead of it (verl#7004). rldoctor does not claim to have solved that.

On verl, add the file logger and point rldoctor at the result — while the run is going, or afterwards:

python -m verl.trainer.main_ppo ... trainer.logger='["console","file"]'
rldoctor diagnose ./<project_name>/<experiment_name>.jsonl

In CI

rldoctor diagnose ./logs/train.jsonl --fail-on critical --format markdown -o $GITHUB_STEP_SUMMARY

What it checks

Every finding carries evidence (the numbers it fired on), a prescription (what to change, with the config key), and references (the paper the threshold comes from). A finding without a suggested change is a complaint, not a diagnosis, and the test suite enforces that every warning has one.

Run rldoctor detectors for the full list, rldoctor fields for the metric aliases.

It reads the log you already have

Detectors are written against a canonical schema; framework keys are resolved onto it in three passes — exact alias, suffix match, then normalised token match. actor/entropy, entropy, policy/entropy and some_new_worker/entropy all resolve to the same field, so a framework renaming a key in a point release degrades gracefully instead of silently producing a worse report. Anything unresolved is reported rather than dropped:

  3 log keys were not recognised: actor/my_custom_metric, ...
  If one of those is a metric rldoctor should understand, please open an issue.

Known-good with verl (W&B, TensorBoard, or its file logger, including val-core/<source>/acc/mean@N validation metrics), TRL (including trainer_state.json from any HuggingFace checkpoint), and any loop that hands us a list of dicts. One required dependency: numpy.

Validated against real runs

Point it at a HuggingFace checkpoint's trainer_state.json — every checkpoint-N/ directory ships one, which makes it the most widely available real training log there is:

rldoctor diagnose ./checkpoint-1200/trainer_state.json

Run against five public GRPO checkpoints (25 to 5,001 logged steps; 1.5s for the largest), it found:

What that exercise actually bought

Three bugs, none of which the simulated corpus could have found, and one capability that only existed because a real run had a shape I had not imagined:

A false positive. The first run reported "4 non-finite gradient norms — training is numerically dead from step 305" on a run whose gradient norms were entirely finite. That log records eval metrics on a different cadence from training metrics, so the aligned series carries nan at every eval-only row — and the detector counted the padding. Every real run with a held-out eval has that shape.

A second false positive. On an 80-step run, six "gradient spikes beyond 8 robust sigma" whose largest value was twice the median. On a tightly clustered series the MAD is tiny, so a statistical outlier need not be a materially large one. A spike now has to be both.

An overclaim. The cost banner announced "~99% of this run's rollout compute produced no learning signal", driven entirely by the truncation rate. But a truncated rollout's advantage is distorted, not zero — that is not the same thing as a degenerate group, which genuinely buys nothing. Truncation no longer feeds the waste estimate; exactly-zero gradient steps, which really do buy nothing, now do.

And a new check. Exactly-zero updates are not a decay and were not something I had thought to look for. Real logs are full of them.

Also fixed while reading the same logs: TRL writes eval_reward with an underscore where verl writes val/test_score with a slash, and handling only the slash form meant eval_rewards/accuracy_reward/mean — a held-out accuracy — was being read as the training reward, which would have inverted the reward-hacking check.

All of it is pinned in tests/test_real_log_shapes.py, and the five runs themselves are re-diagnosed in CI whenever detector code changes (validation/): the raw logs are fetched from their public sources rather than committed, and a change that moves a real-run verdict fails the build. If you point this at a run and it tells you something wrong, that is the most valuable issue you can file.

How it decides

Training curves are noisy, heavy-tailed, and routinely contain a handful of catastrophic steps. Ordinary least squares and plain means over-react to those, which is precisely how a diagnostic earns a reputation for crying wolf. So:

• Theil–Sen slopes and Mann–Kendall significance instead of OLS. A ~29% breakdown point means four exploded steps cannot flip a verdict. (There is a test that asserts OLS is wrecked on the same input.)
• Detrended CUSUM for level shifts, with the noise scale estimated from successive differences. A CUSUM on a rising series otherwise "finds" a shift at the midpoint every time, because a ramp really does have different means either side of any split.
• Model-selecting extrapolation. Entropy decays multiplicatively, so a straight-line projection reads today's steep slope as if it continued forever. On the bundled collapse scenario, asked at step 100 when entropy will hit the floor, a linear-only projection answers step 130 and the log-linear fit answers step 156; the true crossing is step 161. Linear reports half the time you actually have. project_threshold fits both, keeps whichever has the smaller residual in the original units, and tells you which one it used.
• Conditional firing. Entropy decay with reward still climbing is convergence. Entropy decay with reward flat is collapse. Length growth with held-out score growing proportionally is a reasoning model doing its job; length growth without it is padding. The detectors check the second signal before they accuse you of the first.

Thresholds and their rationale are documented in docs/detectors.md. Disagree with one? They are module-level constants with comments explaining the choice.

What it will not do

Being explicit about this, because a diagnostic tool that oversells itself is worse than none:

• It cannot see your rollouts. It reads scalar metrics. It can tell you the shape of reward hacking is present; it cannot show you the exploit. Every reward-hacking prescription starts with "read 20 high-reward rollouts", because that is genuinely the step that finds it.
• Cost figures are estimates. Derived from logged step time, on-demand list GPU prices, and an assumption that ~70% of step wall-clock is rollout generation. Override with --gpu-hour-cost. Treat them as an order of magnitude, which is all they need to be to change a decision.
• Some checks need history. reward_composition reports information rather than a verdict below ~150 logged points, because variance shares during warm-up reflect whichever term is still ramping. It says so rather than guessing.
• It is not a monitoring service. No daemon, no account, no telemetry, no network calls. It reads a log and prints a report.

How it relates to other tools

Design principles

1. A false positive costs more than a false negative. A tool that cries wolf gets uninstalled after one bad alert, and then catches nothing. The healthy-run scenario is asserted across five seeds to produce zero warnings.
2. Never guess from missing data. If a check's inputs are absent it returns SKIPPED and names the key that would have enabled it.
3. Show the arithmetic. Every finding carries the numbers it fired on, so you can disagree with a threshold rather than with the tool.
4. Never break the training loop. LiveMonitor swallows its own exceptions and disables itself with one warning. A monitoring bug must not kill a job that is fine.
5. Stay installable. numpy only. Anything that drags a UI stack or a training framework into a cluster image will not get installed on the cluster where it is needed.

Development

git clone https://github.com/junglezke/rldoctor && cd rldoctor
pip install -e ".[dev]"
pytest                # 143 tests
rldoctor selftest     # detection matrix across all 12 scenarios
ruff check src tests
python tools/make_banner.py   # regenerate the README image
python validation/fetch.py && python validation/run.py --check   # the real public runs

rldoctor selftest is the honest summary of what the tool can and cannot catch — it runs every detector against every simulated pathology and prints the matrix, including the false-positive check on the healthy run.

Contributing

The most valuable contribution is a metric key we do not recognise. If rldoctor diagnose --verbose lists something from your framework, open an issue with the key name and what it measures — that is a one-line fix that makes the tool work for everyone on that stack.

After that: a failure mode you have hit that we do not detect. Adding a detector means one file in src/rldoctor/detectors/ and one scenario in simulate.py. See CONTRIBUTING.md.

Citing

If rldoctor is useful in your work, please cite it — see CITATION.cff.

License

Apache-2.0.