FinConRank: outer-fold adapter collection

Release status: author-approved Apache-2.0 distribution. The author and coauthors have confirmed this licensing decision. See LICENSE, NOTICE, and LICENSING.md for attribution and the scope of this license.

This collection contains ten existing fold-specific epoch-3 LoRA adapters, not a single all-data deployment model. en/outer_0 through en/outer_4 and zh/outer_0 through zh/outer_4 identify the actual outer models in the frozen query-group nested OOF evaluation on a shared fixed document corpus. The historical run name contains "strict report-group"; that name does not establish page or report non-exposure. No inner, T2 sensitivity, best-adapter, or DocReRank checkpoints are substituted. See checkpoint_index.json for hashes.

Loading Relationship

• Base and processor/tokenizer: Qwen/Qwen2-VL-2B-Instruct, revision 895c3a49bc3fa70a340399125c650a463535e71c.
• Training initialization: lightonai/MonoQwen2-VL-v0.1, revision 5d718b52d534290938ed79deedc620a7f6a6254f.
• FinConRank continues updating that LoRA. At inference, load the Qwen base and one FinConRank adapter directly. Do not stack Mono again, merge folds, or average weights.
• LoRA rank/alpha: 32/32. Actual target modules are in each adapter configuration. The released safetensors and adapter configs are byte-identical to the verified original scoring checkpoints; no weights were converted or modified.

Offline Inference

Use an already available copy of the pinned base/processor snapshot. This example never downloads a model. Explicitly select an adapter; there is no default fold or unified root adapter.

CUDA_VISIBLE_DEVICES=0 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 \
python <release>/infer.py \
  --base-model <local-pinned-Qwen-snapshot> \
  --adapter <release>/en/outer_0 \
  --query 'What was the revenue in 2023?' \
  --image <your-page.png> --output <scores.json>

Current implementation requires CUDA. See requirements.txt for the tested environment, not a promise of compatibility with arbitrary newer versions. Install the listed PyTorch CUDA build using your own suitable package source.

financial_constraints.py contains the original deterministic bilingual parser and financial dictionary. It uses only the query, never qrels or scores. Missing information remains unknown. Entity and alias coverage is heuristic. Alternatively, provide your own structured constraints with --constraints-json <file.json>; the schema follows extract_constraints. This option supports replaying original frozen constraints without silently changing prompt contents.

scoring.py preserves the actual prompt, user-only image/text message, base processor chat template with generation prompt, and optimized final-position projection from the original scorer. It returns logit(True) - logit(False); it does not generate an answer, apply a sigmoid, or perform system fusion. FP16 and SDPA are used. Image processor max_pixels is overridden to 602112 (training used 200704). Processor reference configs/template are provided in processor_reference/; loading still uses the explicit pinned base snapshot. Batch size one is used for small-sample validation and this example. Scores need not be bitwise identical across hardware, batch shapes, or software versions.

Scope And Limitations

Evaluation setting: query-group evaluation on a shared, fixed document corpus. Queries are grouped using report identifiers, but document pages are shared and negative sampling can expose pages from other reports. Query-group separation is not page-level or report-level document isolation. These results must not be described as strictly unseen-page or unseen-report generalization. No isolation repair or new evaluation was performed for this release.

This is only the visual reranker component. It does not include ColQwen2 retrieval, Mono candidate scoring, Metric-Aware calibration, fusion, or their learned parameters. Full-system metrics are not individual-adapter metrics, and no such performance claim is attached here. Publication of these weights alone is not end-to-end reproduction or evidence of wholly unseen-page/report generalization. Known cross-report negative-page exposure and all historical audit boundaries must be retained when discussing the original experiments. Choose a checkpoint by its declared fold, not by held-out performance. No all-data model is supplied.

No benchmark images, reports, qrels, training pairs, optimizer states, internal logs, authentication material, or server paths are included in the upload files. User-provided financial documents may have independent rights restrictions.

Attribution And Licensing

Credit Qwen/Qwen2-VL-2B-Instruct and lightonai/MonoQwen2-VL-v0.1 as the base and initial adapter. FinConRank supplies financial-domain adaptation and its associated query-constraint handling; it does not claim authorship of upstream architectures. The author and coauthors approved Apache-2.0 for FinConRank's contributions. Upstream components retain their applicable licenses and attribution obligations; this decision does not relicense third-party material. Historical local resource records identify both upstream models as Apache-2.0, but their exact license and model-card texts were not present in the cached snapshots and were not newly retrieved for this packaging task. See LICENSING.md for that provenance boundary and NOTICE for attribution. The base model itself is not included.

Manual Upload Template

hf upload <OWNER>/<REPOSITORY> <hub_files-directory> . --repo-type model

No repository was created, no login was performed, and no upload was executed.