qwen3-embedding-0.6b-lrat-broad-causal-filter-v1

Qwen3-Embedding-0.6B LRAT Broad Causal Filter v1

This checkpoint is a Qwen3-Embedding-0.6B retriever trained only from the organizer-provided LRAT agent trajectories and their derived query-positive-negative rows.

Method

The starting checkpoint is the compliant broad-g16-v2 continuation of Qwen/Qwen3-Embedding-0.6B. A Qwen3.5-397B auditor inspects each official browse event together with the immediately surrounding official trajectory reasoning and labels whether the browsed document causally advanced the agent's reasoning. Training retains 64,606 high-confidence rows with utility and groundedness at least 2, removes exact query-level positive/negative conflicts, and preserves the original LRAT reasoning-length weights.

The auditor does not generate or import queries, answers, documents, evidence, or leaderboard-derived labels. No A-board query, qrel, per-query outcome, or leaderboard result is used as training supervision.

Training

• Starting checkpoint: qwen3-embedding-0.6b-lrat-broad-g16-v2
• Official-trajectory rows: 64,606
• Global batch size: 256 on 4 GPUs
• Group size: 10
• Epochs / optimizer steps: 1 / 252
• Learning rate: 2e-7 with 0.1 warmup ratio
• Precision: BF16
• Loss: normalized, cross-device weighted contrastive loss
• Pooling: last token, with the source tokenizer's EOS behavior preserved

Local Evaluation

Aggregate dense evaluation on all 830 released A-evaluation queries improves Recall@50 from 0.296163 to 0.300927 and Hit@50 from 0.707229 to 0.721687 against the same-stack broad-v2 control.

On two fixed-seed Agent splits (queries 1-150 and 151-300), the conservative 300-query aggregate is:

One candidate query hit a reproducible evaluator-side vLLM parsing error on all three retries. The table conservatively imputes it as zero recall, zero success, and the full 20-step budget. Relative to the submitted front-success-v1 checkpoint under the same two splits, Total improves by 0.8671, Success by 2.6662 points, and average search steps by 0.2400, while Recall decreases by 0.7385 points.

These local measurements are model-selection evidence and are not official leaderboard results.