PivotRL-Correct-collection

Turn-level GRPO training data for Qwen3.5-35B-A3B-Correct-GRPO: 4,582 stop-decision pivot states (4,573 distinct questions) harvested from web-search agent trajectories of the Correct SFT policy — the correctness-filtered arm of the four-arm study. Each row is a serialized agent-conversation prefix ending at a decision point between stopping to answer and continuing to search, with gold answer keysets for a deterministic stop-verifier reward. It is the Correct-base counterpart of PivotRL-Asis-collection, built with the same protocol so the two RL runs are comparable arm-to-arm.

Construction

Pivot states were selected by K=8 rollout profiling: from each candidate mid-trajectory state, 8 continuations were rolled out with the (Correct) SFT policy and scored by the deterministic stop-verifier; only mixed-outcome states (profiled success rate 0 < μ̂ < 1; observed range 0.125–0.875) were kept, so every state carries learning signal under group-relative advantage. Hard states (μ̂ < 0.25) additionally carry a conditional-seeding flag (allow_seed) permitting evidence-consistent hinting during RL rollouts. The set comprises 4,432 grounded finalize-decision states plus a 150-state out-of-distribution first-turn control complement that regularizes against degenerate always-stop behavior.

Questions originate from an in-house web-search QA corpus (English and Korean); trajectories were generated by the SFT policy itself against a live browsing tool surface. No benchmark data was used at any stage (see Contamination below).

Fields

Field	Type	Description
id	str	unique pivot id (stoppivot_<traj>_<turn>)
prompt	str (JSON)	serialized message list: browsing system prompt, user question, then alternating assistant / tool turns up to the pivot
tools	str (JSON)	function specs of the browsing tool surface: search, open, find
label	str (JSON)	{"keys": [[...], ...]} — gold answer keysets; each inner list is a set of acceptable surface forms, all groups must be matched for a finalize to be verified correct
metadata	str (JSON)	per-state annotations (below)

metadata keys: evidence_state (grounded/ungrounded), a_star (reference action: finalize/toolcall), a_star_text (reference finalize text with evidence-citation markers), sft_action_name (action the SFT policy took at this state in the source trajectory), trajectory_id, assistant_depth (prefix depth in assistant turns), mu_profiled (μ̂ from K=8 rollout profiling), allow_seed (conditional-seeding flag), key_cursors (cursors of evidence pages containing the gold keys; grounded rows), track (A = grounded finalize states; C_* = control complement), plus auxiliary source on control rows.

from datasets import load_dataset
import json
ds = load_dataset("Minbyul/PivotRL-Correct-collection", split="train")
row = ds[0]
messages = json.loads(row["prompt"]); label = json.loads(row["label"]); meta = json.loads(row["metadata"])

Contamination screening

Both this file and its Asis-arm counterpart were screened row-by-row against three browsing benchmarks — LiveBrowseComp (335 questions), K-BrowseComp (300), and BrowseComp-EN (1,266) — by matching each row's first user message with three matchers: exact normalized equality, 120-char prefix containment, and 8-gram Jaccard ≥ 0.5. Result: 0 hits on all benchmarks and all matchers (rescan dated 2026-08-20; consistent with the corpus-wide screen of 2026-08-13). A field-level scrub scan additionally verified that no internal corpus identifiers, infrastructure paths, or generator-model references appear in system/user/instruction or metadata fields; model names occurring inside retrieved web text are genuine web content.

Associated models

• Trained on this data: Qwen3.5-35B-A3B-Correct-GRPO (turn-level GRPO, K=8, deterministic stop-verifier reward, KL 0.01 to the SFT reference; released checkpoint at ~14 effective steps — this base's earlier safe window)
• SFT policy / reference: Qwen3.5-35B-A3B-Correct
• Asis-arm counterpart dataset: PivotRL-Asis-collection

Intended use & limitations

Research artifact for studying stop-decision RL, over-search reduction, and base-dependence of the safe RL window in tool-using web agents. Prompts are tied to a specific browsing tool surface and citation format; reward semantics assume the deterministic stop-verifier described in the model cards. Tool-turn contents are snapshots of live web pages at generation time and may contain outdated or incorrect claims; they are included as environment observations, not as ground truth.