Configs

prime-rl uses pydantic-config — a Pydantic-based TOML + CLI config system (no tyro). Every entrypoint accepts TOML files via @ and CLI overrides.

Loading and composition

uv run rl @ examples/basic/reverse-text/rl.toml                                  # single TOML
uv run rl @ examples/basic/reverse-text/rl.toml --max-steps 50                   # CLI override
uv run rl @ base.toml @ overlay.toml                                       # left-to-right merge
uv run rl --model @ model.toml --data @ data.toml                          # nested section files
uv run rl @ base.toml --trainer @ trainer.toml --trainer.lr 1e-3           # mixed

Resolution order: CLI > config files (left-to-right) > class defaults. Merging is deep — unset fields in an overlay are preserved from the base. output_dir has one extra fallback: CLI > config files > $PRL_OUTPUT_DIR > "outputs".

Naming: CLI uses kebab-case (--vllm.max-model-len); TOML uses snake_case (max_model_len).

Inspect & validate

uv run rl --help                                  # all fields and defaults
uv run rl @ rl.toml --dry-run --output-dir /tmp/x --run.name check # write resolved JSON to /tmp/x/check/configs/latest

Each attempt also writes configs/attempt_<n>/command.txt. It records the shell-safe launch command, including CLI overrides. configs/latest points to the current attempt.

Validators

Incompatible combinations (e.g. CP requires flash attention) must raise in a model_validator at resolve time, not at runtime. When renaming a field, remove the old spelling: no validation_alias, no auto-translating mode="before" validator. The old key then fails as an unknown key, which is the signal. An alias that stays forever is worse than a break — it never gets retired, and a key whose meaning changed silently misconfigures the run.

Special syntax

No inline tables — checked-in configs use [section] headers or dotted keys, never key = { ... }.

Sources are one block — inside a [[...source]] entry, write nested sub-configs as dotted keys in the same block (env.taskset.id = "...", env.agent.harness.id = "..."), not one subsection header per nested config. Nested arrays of tables (e.g. [[orchestrator.train.source.env.taskset.task.judges]]) keep full-path headers — they attach to the preceding [[...source]] entry.

Booleans — CLI --flag / --no-flag; TOML must be explicit (enforce_eager = true).

None — TOML has no null, use the string "None" (max_model_len = "None"); CLI: --vllm.max-model-len None.

Lists — TOML uses array of tables; later config files replace lists wholesale, so overlays must include the full desired list:

[[orchestrator.train.source]]
name = "reverse-text"
env.taskset.id = "reverse-text"
env.agent.harness.id = "null"
env.agent.runtime.type = "subprocess"

[[orchestrator.eval.source]]
name = "reverse-text-eval"
env.taskset.id = "reverse-text"
env.taskset.split = "test"
env.agent.harness.id = "null"
env.agent.runtime.type = "subprocess"

CLI: --orchestrator.train.source.0.env.taskset.id reverse-text or --orchestrator.eval.source.0.env.taskset.id reverse-text.

The sft entrypoint takes the same eval shape at the top level for online evals: [eval] + [[eval.source]] (with [inference] for the server), e.g. --eval.source.0.env.taskset.id reverse-text.

Dicts — TOML uses a section; CLI takes a JSON string: --trainer.env-vars '{"key1": "value1"}'. This works for plain dict fields only — nested pydantic-model fields (e.g. algo) reject JSON strings; use dotted keys (--orchestrator.algo.type max_rl) or a TOML overlay file.

vLLM pass-through — [inference.vllm] uses vLLM's own argument names (model, tensor_parallel_size, data_parallel_size, max_model_len, ...) and forwards any key to the vLLM server, typed by prime-rl or not: [inference.vllm] max_num_seqs = 256, or --inference.vllm.max-num-seqs 256 on the CLI. CLI values are JSON-coerced, so dict-valued vLLM args work as --inference.vllm.compilation-config '{"cudagraph_mode": "NONE"}'. Non-vLLM knobs (router, deployment, weight broadcast, kv-cache offload, env vars) stay on [inference] itself.

Discriminated unions — set the type field to pick the variant ([orchestrator.algo] type = "max_rl"). Omit type to keep the default variant.

RL loss — [trainer.loss] defaults to IPO with eps = 0.1, adv_tau = 1.0, and kl_tau = 1e-3. Omit the section to use these defaults. Set type = "custom" with import_path and optional kwargs to load a custom RL loss. The ce and ref_kl components are fixed.

Algorithms — [orchestrator.algo] type = "grpo" | "max_rl" | "rae" | "hierarchical_grpo" | "opd" | "opsd" | "sft" | "echo" — the type names the algorithm (credit assignment + loss routing, fused), and each type's class defaults are its vetted setting; any other key you set is your own assembly (e.g. [orchestrator.algo.roles.user] alpha = 0.1 for echo — setting any echo role replaces the whole role table). hierarchical_grpo is only valid with a proposer-solver env: it compares solvers with attempts on the same proposed problem and proposers with other proposals in the group. There is no preset layer, and no config hook that points at user code — a new algorithm is a named class in the repo (subclass Algorithm, register it). Per-source override: [orchestrator.train.source.algo] type = "opd" (the source assembles its own algorithm). prime-rl only hosts the trainable policy; frozen models are inline external endpoints on the algorithm, named where the model is used — [orchestrator.algo.teacher] for opd (the frozen model scored against), [orchestrator.algo.sampling.source] for sft (the model it samples from), each with name + base_url. There is no shared teacher slot. opsd declares no model — it self-distills against the live policy. See docs/algorithms.md.

BaseModel | None fields — bare flag enables defaults; nested override enables and sets:

--model.compile             # enables compile with defaults
--model.compile.fullgraph   # enables and sets fullgraph=true

In TOML, an empty section header ([ckpt]) does the same.

RL trainer token exports

For rollout debugging, enable trainer-side token export with trainer.enable_token_export = true (or --enable-token-export when running the trainer entrypoint directly). It writes one JSONL record per exported sequence under <run_dir>/token_exports/step_<step>/rank_<rank>.jsonl. Each record stores aligned per-token arrays for token ids, loss mask, component weight streams (rl/ce/ref_kl), advantages, entropy, mismatch KL, inference/trainer logprobs, importance ratios, probability deltas, and masking diagnostics. It does not decode token text in the trainer.

enable_token_export = true

Leave it unset for normal training. When enabled, it exports every sequence from each exporting rank.

Key files

• packages/prime-rl-configs/src/prime_rl/ — config classes under configs/; utils/config.py re-exports BaseConfig and cli
• configs/debug/ — minimal debug configs
• examples/ — full example configs