[faresobeid]

_parse_compact_logprobs gets a second row mode. With "samples": true in the compact payload, each row is the sampled token followed by K i.i.d. draws from the sampler distribution q. The client returns the draws as completion_top_ids / completion_top_logprobs in draw order, repeats kept. Without the flag, parsing is unchanged: the head is ranked, deduplicated and tie-trimmed.

prime-rl uses these rows for Monte Carlo score centering without sampling replay: E_q[w s] ≈ mean_k w(v_k) s(v_k). Each draw has weight 1/K, so repeats must survive parsing.

Row checks: the first entry must be the sampled token, and there must be at least one draw. The existing value and offset validation still applies. Sampling-mask parsing is skipped for these rows (they carry no replay support).

Validation

uv run pytest tests/test_client.py (105 passed, 6 skipped), ruff check, ruff format --check; CI green (incl. Ty).

Dependencies

PrimeIntellect-ai/prime-rl#4232 emits these rows and pins deps/renderers to this branch (a727c5e). Re-pin to the merged commit when this lands.

Note

Support sampler-draw repeats in compact logprob parsing in client.py

• Adds a samples flag to _parse_compact_logprobs in renderers/client.py. When true, the parser skips sampling-mask parsing and keeps every packed draw after the sampled-token entry, including duplicate token IDs, in original order.
• Sampler rows are rejected when the first packed ID is not the sampled token or when no additional draw exists. The candidate-selection path is unchanged when samples is false or absent.
• Adds a regression test in tests/test_client.py with repeated token IDs to verify repeats are preserved.

^{Macroscope summarized a727c5e.}