TaxFlow V8.1 Evidence Node LoRA
TaxFlow V8.1 is a competition/research LoRA adapter for Qwen3-VL-2B-Instruct.
It is trained for a narrow accounting-firm workflow task: composing conservative, reviewable tax workpaper evidence nodes from upstream OCR/PDF text-layer output, official tax-form template context, Evidence RAG retrieval, page/cell metadata, and quality gates.
This adapter is not a general OCR model, not a tax-law reasoning model, and not a manager-review replacement.
What Was Trained
The adapter was trained to make the model behave more like a TaxFlow workpaper component:
- preserve
form_code, field_id, source page/cell, and visible value context;
- keep official instruction / RAG source metadata near the field;
- output reviewable Evidence Node JSON instead of free-form prose;
- mark OCR/template/RAG conflicts as
conflict_requires_review;
- provide DataSnipper-style handoff fields when source anchors are available;
- avoid pushing uncertain values directly into O SWP workpapers.
Training Run
- Hardware: single RTX 4090 24GB training run
- Base model:
Qwen/Qwen3-VL-2B-Instruct
- Fine-tuning: LoRA adapter
- Final checkpoint:
checkpoint-342
- Final eval loss:
0.1521
- Adapter files exported at repository root
Small-Sample Fair Compare
This is a small smoke benchmark, not a statistically significant production benchmark.
| Split | Model | N | Readiness | Parse | Value | Stress Reject | Conflict Status | RAG Source | DataSnipper |
|---|
| mixed | base | 8 | 43.8% | 87.5% | 87.5% | 87.5% | 87.5% | 0.0% | 0.0% |
| mixed | taxflow | 8 | 35.4% | 50.0% | 50.0% | 50.0% | 50.0% | 50.0% | 0.0% |
| stress | base | 12 | 47.2% | 100.0% | 100.0% | 0.0% | 0.0% | 0.0% | 0.0% |
| stress | taxflow | 12 | 61.8% | 66.7% | 66.7% | 66.7% | 66.7% | 66.7% | 66.7% |
| showcase | base | 8 | 47.9% | 100.0% | 87.5% | 100.0% | 100.0% | 0.0% | 0.0% |
| showcase | taxflow | 8 | 47.9% | 62.5% | 62.5% | 62.5% | 62.5% | 62.5% | 12.5% |
Safe Claim
TaxFlow V8.1 shows a narrow but defensible advantage in targeted production-stress conflict cases: compared with the base model, it is more likely to return review-required Evidence Nodes with RAG source context and DataSnipper handoff fields.
Explicit Non-Claims
Do not claim from this run that:
- TaxFlow is generally better than the base model across all extraction tasks;
- the LoRA improves raw OCR visual acuity;
- the model is production-ready on 11th-gen CPU office laptops;
- the model can replace human reviewer judgment;
- the benchmark is statistically significant.
Mixed/showcase samples did not show a general win. The next iteration should shorten the schema, use constrained decoding / JSON repair, and run a larger benchmark.
Files
adapter_model.safetensors: LoRA adapter weightsadapter_config.json: PEFT adapter configTRAINING_SUMMARY_PUBLIC.json: sanitized training metadataeval/: score JSON, dashboard, and fair-compare result archivedocs/HF_MODEL_CARD_V8_1_RESULT.md: detailed release note
Public Release Hygiene
This public-sanitized package intentionally excludes raw training logs, raw prediction archives, local machine paths, and base tokenizer files. Use the base model tokenizer/processor from Qwen/Qwen3-VL-2B-Instruct.
