model stringclasses 1
value | layers listlengths 4 4 | hidden_size int64 4.1k 4.1k | rows_sha256 stringclasses 4
values | n_rows int64 26.2k 26.2k | rows_file stringclasses 1
value | feature_schema stringclasses 2
values | method stringclasses 1
value | anchor stringclasses 2
values | complete bool 1
class | limit int64 0 0 |
|---|---|---|---|---|---|---|---|---|---|---|
llava-hf/llava-1.5-7b-hf | [
16,
20,
24,
28
] | 4,096 | 98bc51800bc241518d05029ff23eeea6eb83be7cea686d29c4e463c3f72f4342 | 26,150 | train.jsonl | prior_object_last_token_v1 | MLP-Probe | final prompt token before the yes/no answer | true | 0 |
llava-hf/llava-1.5-7b-hf | [
16,
20,
24,
28
] | 4,096 | 2d53d25eadc9c7392f9b2ee8970c9487d69b346f55fab16da1d8acb4f8ef0d15 | 26,150 | train.jsonl | prior_object_last_token_v1 | MLP-Probe | final prompt token before the yes/no answer | true | 0 |
llava-hf/llava-1.5-7b-hf | [
16,
20,
24,
28
] | 4,096 | 423ca294265715dcac288569daafb7d3f8cf6034b7c005caf24daf0f64b6eeeb | 26,150 | train.jsonl | prior_object_object_token_v1 | MLP-Probe | last subtoken of the object name in the prompt | true | 0 |
llava-hf/llava-1.5-7b-hf | [
16,
20,
24,
28
] | 4,096 | 9ea79257bb97fcca7d20c4599e9a5d424f6252adbde31e4db8544e96fb1b6987 | 26,150 | train.jsonl | prior_object_object_token_v1 | MLP-Probe | last subtoken of the object name in the prompt | true | 0 |
Isolated question vs generation state: analysis outputs
Analyses of object-hallucination probes on LLaVA-1.5-7B comparing the generation state of the model's own
caption (Normal) with an isolated single-object yes/no question (Isolated). Each folder holds the report,
per-object scores, figures and a run_manifest.json with the dataset, design, code commit and command.
| Folder | Observation | Content |
|---|---|---|
position/coco_llava15_7b |
O1/O2 | AUROC by distinct-object position, four probes, Normal vs Isolated, COCO holdout |
position/amber_llava15_7b |
O1/O2 | the same on AMBER test |
mask_object/coco_llava15_7b |
O4 | object-mask test of image sensitivity, four probes, both settings |
prefix_context/coco_llava15_7b |
O5 | text placed before the isolated question, frozen probes |
Detection records, checkpoints and the prior x object ablation are in
ToiTenBao/hallucination-evaluation-results. The living summary of all observations is
docs/experiments/FINDINGS_isolated_vs_generation.md on branch exp/prior-object-probe-ablation of the code
repository. Seeds 0, 42, 1337; metrics are seed means, never ensembles.
- Downloads last month
- 1,117