Family
Qwen2.5-VL-7B + LoRA
shadowApache 2.0multimodalboth
Document VLM.
Vision-language model for document understanding (OCR + extraction) — Qwen2.5-VL-7B with LoRA fine-tuning
Foundation · Qwen2.5-VL-7B + LoRA
What it does
The honest description.
Vision-language model for document understanding (OCR + extraction) — Qwen2.5-VL-7B with LoRA fine-tuning
Architecture
How it works.
PDF documents (2K synthetic + real bank statements, salary slips, ITR, insurance, loan statements, receipts, invoices). Encoded to images via vLLM multimodal pipeline. Structured JSON field extraction per document type (issuer, date, amount, account holder, etc.) with per-bank F1 scores. Target > 90% F1.
Training data & drift
What it learned from, and how we watch it.
Source
Inputs: PDF documents (2K synthetic + real bank statements, salary slips, ITR, insurance, loan statements, receipts, invoices). Encoded to images via vLLM multimodal pipeline.
Drift baseline · current quarter
Drift baseline not yet documented for this model.
What it's NOT
- Qwen2.5-VL-7B + LoRA — open-weight Apache-2.0 Qwen/Qwen2.5-VL-7B-Instruct (NOT a proprietary qwen-vl-max Bailian variant).
- vLLM-served.
- Track-4 PR-B registered DocumentVLMTrainer + wired real-KYC parquet dispatch (kyc.parquet via MODEL_TO_UPLOAD_TYPE), DORMANT behind DOCUMENT_VLM_TRAINING_ENABLED (default False) — a KYC upload writes the parquet but NEVER auto-trains.
- TEXT corpus-gate (UIDAI/PII redaction + minor-exclusion + DSR §13 + post-redaction leak re-scan) fires BEFORE the collator (boot guard document_vlm_training_applies_compliance_gate_before_collator).
- IMAGE-PIXEL UIDAI redaction NOT YET IMPLEMENTED — real-KYC training FENCED: _train_with_transformers REFUSES on any raw image reference (fence runs before collator).
- trainable stays FALSE (GPU-gated).
- No auto-promote.
Who calls this model
Across the house and the API.
Internal routes
No public routes documented yet.
Version history · immutable once published
Every release, traceable.
Data Mukodi holds about you is subject to DPDP §12 erasure (use /consent in the app), and any concern about this model can be raised via [email protected] under DPDP §13 grievance, or via /security.