Skip to content

Latest commit

 

History

History
80 lines (64 loc) · 3.44 KB

File metadata and controls

80 lines (64 loc) · 3.44 KB

Model Card: E6 Grounded VQA Adapter and E19 Decoding System

Updated: 2026-08-22

Model description

  • Base: Salesforce InstructBLIP Flan-T5-XL.
  • Adapter: E6 LLM LoRA, rank 8, alpha 16, dropout 0.05.
  • Trainable parameters: 4,718,592 (about 0.201% of the model).
  • Training data: a deterministic random 10,000-example subset of VQAv2 train, one epoch.
  • Final VQA system: the unchanged E6 adapter plus E19 short-answer reranking for routed yes/no questions.

E19 is a decoding component, not a separately trained checkpoint. External POPE, CHAIR, and HallusionBench results use direct deterministic generation unless the result explicitly says pair-logprob.

Intended use

This research model is intended for short-answer visual question answering, controlled PEFT experiments, and study of visual-language alignment and hallucination trade-offs. It is not intended for safety-critical decisions or unverified real-world factual claims.

Main results

Evaluation Result
Full VQAv2, E6 direct 70.609
Full VQAv2, E6 + E19 71.620
VQAv2-derived MCQ 85.12
POPE macro accuracy, E6 84.933
POPE macro accuracy, zero-shot 83.944
CHAIRs, E6 / zero-shot (lower is better) 43.40 / 31.80
CHAIRi, E6 / zero-shot (lower is better) 14.53 / 10.93
CHAIR object recall, E6 / zero-shot 66.98 / 62.16
HallusionBench question accuracy, E6 / zero-shot 52.70 / 54.30

E6 significantly improves POPE accuracy by 0.989 points, but does so by raising object recall and affirmative answers while lowering precision. On open-form CHAIR, E6 significantly raises both object recall and hallucination. On HallusionBench, its −1.594 point accuracy change is not significant and its false-positive rate increases. The supported conclusion is a measurable recall–hallucination trade-off, not universal hallucination mitigation.

Alignment evidence

On the fixed VQAv2 diagnostic subset, replacing the correct image with a mismatched image lowers accuracy by 34.75 points and changes 65.9% of answers. This establishes that the model uses visual input. Mismatch, complementary, hard-pair, Q-Former, and grounding objectives did not consistently improve both VQA and alignment, and are retained as controlled negative or trade-off results.

Limitations

  • Only the E6/E19 system has a complete 214,354-question VQAv2 validation run; many training ablations use a fixed 1,000- or 5,000-question development set.
  • Most training ablations have one seed. E14 is the main three-seed exception.
  • CHAIR uses a documented Python 3 port of the original object rules with a deterministic tokenizer instead of the original Python 2 pattern/nltk stack.
  • HallusionBench text-only rows use a white image because InstructBLIP requires an image input.
  • Full official THRONE, human evaluation, and closed-model comparisons have not been run.
  • The VQAv2-derived MCQ set has automatically sampled distractors and is not an official multiple-choice leaderboard.

Reproducibility artifacts

Server project root: /root/autodl-tmp/vision-language

  • Adapter: outputs/E6_instructblip_llm_lora_r8_train10k_random/final-adapter
  • Full direct VQA: outputs/E15_E6_instructblip_full_vqav2_val214354
  • E19 full VQA: outputs/E19_E6_short_answer_rerank_full_val
  • POPE: outputs/H1_E6_official_COCO_POPE
  • CHAIR: outputs/H2_E6_CHAIR500
  • HallusionBench: outputs/H3_E6_HallusionBench
  • Consolidated report: reports/FINAL_EXPERIMENT_SUMMARY.md