letter-counting-legacy-rl

Legacy letter-counting RL protocol: first 64 generated tasks train, next 64 held out.

Field	Value
Reference	nodus/letter-counting-legacy-rl@1.0.0
Image	nodus/env-letter-counting-legacy-rl:1.0.0
Publisher	Open-Thought
Category	Reasoning
Readiness	Research
Modes	Train, Evaluate
Reward	Scalar
Held-out measures	TrainedTask
Splits	64 train, 64 test (disjoint by canonical identity)
Licenses	code Apache-2.0, data Apache-2.0
Source	https://github.com/open-thought/reasoning-gym

Grading

Completions are graded by the platform, never by the trainer: the trainer submits {taskId, completion} batches and the verdicts come back as task events.

Grader	Kind	What it checks
score-answer	Program	Raw completion scored by pinned Reasoning Gym; only full credit is Correct

Sample task

One line of nodus-env tasks --split test --seed 0; tasks never carry answers.

{
  "metadata": {
    "family": "letter_counting",
    "systemPrompt": "Reply with the final answer only. Do not restate the question, do not show working, and do not add any words, labels or punctuation around it."
  },
  "prompt": "How many times does the letter \"t\" appear in the text: \"are far more complex than all that In real life every act\"?",
  "taskId": "letter-counting-legacy-rl:test:0:0"
}

Examples

Each example is a TrainingJob template. A baseline is shown only where the example was measured by running it.

Example	Mode	Runtime	Model	Tasks	Baseline	Trained	Measured on
letter-counting	Train	nodus/grpo-lora	Qwen/Qwen3-1.7B @ 70d244c	64	not measured	not measured	not measured

Run one with a server dry-run first:

$ nodus create trainingjob my-run --from-example nodus/letter-counting-legacy-rl:letter-counting --dry-run=server -o estimate

import nodus

job = nodus.recipes.TrainingJob.from_example("nodus/letter-counting-legacy-rl:letter-counting")
plan = job.preview()
run = plan.run(max_cost=5)
print(run.wait().summary)