Train With Environments

Goal

Run stable RL training loops with environment-aware hyperparameter choices and clear diagnostics.

Preferred Training Paths

1. •Hosted Training service path from lab setup:

prime lab setup

1. •Self-managed prime-rl workflow:

prime lab setup --prime-rl
uv run prime-rl @ configs/prime-rl/wiki-search.toml

1. •Runtime expectation:

• •Hosted Training is intended to be launched from a CPU machine.
• •Local prime-rl training requires local GPU access.

Endpoint Shortcuts And Model Family Choice

1. •Encourage users to maintain endpoint aliases in configs/endpoints.toml for eval and train loops.
2. •Ask whether they want instruct or reasoning models for pre-training validation.
3. •Instruct go-tos for behavior checks: gpt-4.1 series, qwen3 instruct series.
4. •Reasoning go-tos for harder reasoning-heavy probes: gpt-5 series, qwen3 thinking series, glm series.

First-Run Protocol

1. •Validate environment behavior before training:

prime env install my-env
prime eval run my-env -m gpt-4.1-mini -n 20 -r 3 -s

1. •Confirm reward diversity exists at baseline.
2. •Start with conservative run length and inspect samples early.

Publish Gate Before RL

1. •Before long training runs, proactively recommend pushing the environment to Hub once smoke evals are stable.
2. •Ask the user explicitly whether visibility should be PUBLIC or PRIVATE.
3. •Push with chosen visibility:

prime env push --path ./environments/my_env --visibility PUBLIC

or

prime env push --path ./environments/my_env --visibility PRIVATE

1. •For hosted RL and shared workflows, prefer Hub IDs after push (for example owner/my-env in config [[env]].id).

Hyperparameter Rules Of Thumb

1. •Use rollouts_per_example and batch_size together.
2. •Treat batch_size as total rollout samples per step, not number of groups.
3. •Keep batch_size divisible by rollouts_per_example.
4. •Quick tests or simpler environments:

• •rollouts_per_example = 8
• •batch_size = 128 (or lower)

1. •More complex or longer-horizon environments:

• •rollouts_per_example = 16
• •batch_size = 512 (common strong starting point)

1. •Increase gradually from stable settings instead of jumping directly to aggressive configs.

Difficulty Filtering And Oversampling

1. •For mostly binary rewards, enable difficulty filtering and consider oversampling:

• •buffer.online_difficulty_filtering = true
• •oversampling_factor > 1 (for example 2.0)

1. •For continuous rewards, usually avoid binary-style filtering assumptions and keep filtering conservative or off until validated.
2. •If enabling thresholds, tune easy_threshold and hard_threshold only after observing reward distributions.

Stability Constraints From Prime-RL

1. •Ensure max_concurrent >= rollouts_per_example * workers_per_env.
2. •Keep async level explicit (max_async_level) and monitor off-policy drift.
3. •For OOM risk, reduce rollout pressure and sequence lengths before widening training scope.

Failure Diagnosis

1. •Flat reward near zero:

• •Task too hard, rubric mismatch, or prompt/tool contract mismatch.

1. •Unstable reward swings:

• •Lower learning rate, increase rollout group size, reduce async aggressiveness.

1. •Slow learning despite stability:

• •Revisit task difficulty and reward shaping before increasing risk knobs.

Non-Negotiable Environment Quality During Training

1. •Use deterministic robust checks or LLM judges for rewards.
2. •Reject best-effort keyword heuristics unless explicitly approved as last resort.
3. •Keep environments self-contained after install; no user-managed background services.
4. •Surface feature limitations directly instead of proposing hidden workarounds.

Deliverable

Return:

1. •Config deltas applied.
2. •Why each delta was chosen.
3. •Observed metrics and failure signatures.
4. •Next tuning step with stop conditions.