This is a Qwen3-Embedding-0.6B compatible dense retriever checkpoint for
XIR competition 1170.
Method
The checkpoint starts from the compliant LRAT broad-g16-v2 retriever and is
trained on 91,324 organizer-derived rows:
64,606 stable trajectory rehearsal rows;
26,718 new query-document edges discovered and independently verified by
three strong models over organizer-provided trajectory text.
The new edges use bounded reliability calibration based on signals fixed
before evaluation: independent-model vote fraction, minimum confidence,
minimum relevance, task-query versus trajectory-query role, and within-graph
document centrality. Calibration changes only relative loss weight within the
augmentation set; its mean weight remains 1.0.
Data Boundary
No external query, answer, passage, or document is used.
No generated text is added to training.
Strong models only relabel organizer-provided trajectories and text.
No leaderboard query, qrel, hidden failure, or per-query result is used as
training supervision.
Training
Base: Qwen/Qwen3-Embedding-0.6B through the compliant LRAT lineage
Global batch: 256
Group size: 10
Learning rate: 2e-7
Epochs: 1
Precision: BF16
Weighted cross-device InfoNCE
Local Evaluation
On the complete 830-query fixed local evaluation:
Dense Recall@50: 0.313709
Dense Hit@50: 0.740964
Fixed-seed Agent recall: 50.431855%
Fixed-seed Agent success: 26.867470%
Average Search calls: 22.944578
Local Agent totals are for paired checkpoint selection and are not claimed to
equal the organizer leaderboard score.
