Publish AIST-87M human-memory embedding model

Browse files

Files changed (7) hide show

.gitattributes +1 -34
AIST-87M.safetensors +3 -0
README.md +155 -0
aist87m_memory_slice_release_report.json +1275 -0
aist87m_memory_slice_release_report.md +85 -0
aist_81m_raw_mn20_lora.yaml +55 -0
parameter_breakdown.json +9 -0

.gitattributes CHANGED Viewed

@@ -1,35 +1,2 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text


























1	*.safetensors filter=lfs diff=lfs merge=lfs -text
2	+ *.gguf filter=lfs diff=lfs merge=lfs -text

AIST-87M.safetensors ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:c93839ee0875e75b4dbb91ff510bb736122d666631f50902f6e587f158ebd7ec
+size 348855664

README.md ADDED Viewed

	@@ -0,0 +1,155 @@

+---
+language:
+- en
+license: apache-2.0
+tags:
+- multimodal
+- embedding
+- trimodal
+- retrieval
+- image-text-audio
+- audio
+- speech
+- memory-augmentation
+- feature-extraction
+library_name: pytorch
+pipeline_tag: feature-extraction
+datasets:
+- custom
+---
+# AIST-87M
+`AIST-87M` is a compact audio + image + speech + text embedding model for
+human-memory augmentation workloads.
+It is the single-audio evolution of the earlier dual-audio tower line: the
+runtime audio path uses one merged native `mn20_as` EfficientAT encoder instead
+of a separate EfficientAT + Whisper dual branch. The LoRA training weights are
+merged into the native audio encoder in this release artifact, so there is no
+separate LoRA pass at inference time.
+Core stack:
+- text: `MongoDB/mdbr-leaf-ir`
+- image: `mobilenetv4_conv_medium.e180_r384_in12k`
+- audio: native merged `mn20_as` EfficientAT encoder
+- projection output: `1280d`
+- Matryoshka slices: `[1280, 768, 512, 256, 128]`
+- exact loaded params: `87,118,774`
+The canonical name follows the Augmem naming standard:
+- `AIST` = audio + image + speech + text
+- `87M` = exact loaded parameter count rounded to integer millions
+## Runtime Contract
+This model returns L2-normalized embeddings in a shared 1280-dimensional space.
+For smaller runtime profiles, truncate to a Matryoshka slice and renormalize:
+```text
+z1280 = l2norm(model(input))
+z768  = l2norm(z1280[0:768])
+z512  = l2norm(z1280[0:512])
+```
+The release safetensors file is self-contained and includes the text encoder,
+image encoder, merged native audio encoder, and the three projection heads.
+## Evaluation Scope
+This release uses a human-memory evaluation slice rather than a broad
+leaderboard sweep. The slice is chosen to match practical memory augmentation
+surfaces:
+- text continuity: duplicate-question and semantic textual similarity tasks
+- image recall: Flickr30k text-image and image-text retrieval
+- audio recall: speech/general-audio text-audio retrieval tasks
+Primary metrics:
+- text continuity: `main_score`
+- image recall: `NDCG@10`
+- audio recall: `NDCG@10`
+## Human-Memory Slice
+Source: `aist87m_memory_slice_release_report.md` and
+`aist87m_memory_slice_release_report.json`.
+| Dim | Tasks | Text continuity | Image recall | Audio recall | Overall |
+|---:|---:|---:|---:|---:|---:|
+| 1280 | 8 / 8 | 0.763 | 0.425 | 0.104 | 0.349 |
+| 768 | 8 / 8 | 0.762 | 0.424 | 0.104 | 0.349 |
+| 512 | 8 / 8 | 0.762 | 0.424 | 0.104 | 0.349 |
+Selected 1280d task scores:
+| Task | Family | Metric | Score | R@1 | R@10 |
+|---|---|---|---:|---:|---:|
+| SprintDuplicateQuestions | Text continuity | main_score | 0.875 | - | - |
+| STSBenchmark | Text continuity | main_score | 0.651 | - | - |
+| Flickr30kT2IRetrieval | Image recall | NDCG@10 | 0.469 | 0.296 | 0.672 |
+| Flickr30kI2TRetrieval | Image recall | NDCG@10 | 0.381 | 0.082 | 0.407 |
+| CommonVoiceMini21T2ARetrieval | Audio recall | NDCG@10 | 0.028 | 0.006 | 0.062 |
+| MACST2ARetrieval | Audio recall | NDCG@10 | 0.110 | 0.033 | 0.214 |
+| UrbanSound8KT2ARetrieval | Audio recall | NDCG@10 | 0.009 | 0.002 | 0.018 |
+| ClothoT2ARetrieval | Audio recall | NDCG@10 | 0.269 | 0.128 | 0.443 |
+## Task-Aligned Comparisons
+Comparisons below are only for locally available, task-aligned runs.
+| Comparison | Dim | Paired tasks | Read |
+|---|---:|---:|---|
+| vs `ES-AIST-81M` | 768 | 8 | lower text continuity; stronger Flickr and selected audio recall |
+| vs native `mn20_as` audio baseline | 768 | 4 | slightly lower selected audio recall on average; UrbanSound8K is flat |
+| vs dual-audio tower | 768 | 6 | smaller single-audio runtime, but lower paired text/image/audio scores |
+| vs `AIST-95M` | 1280 | 2 | only paired Flickr tasks are available locally; `AIST-95M` remains stronger on that pair |
+This release is not presented as a generic MTEB/MIEB/MAEB leaderboard model.
+Broad diagnostic runs contain many task families that are not part of this
+release gate.
+## Architecture
+```text
+Text   -> mdbr-leaf-ir (768-d) -----------------------> DeepProjectionHead-d2 -> 1280
+Image  -> MobileNetV4-Medium (1280-d) ----------------> DeepProjectionHead-d2 -> 1280
+Audio  -> merged native EfficientAT mn20_as (1280-d) -> DeepProjectionHead-d2 -> 1280
+```
+The audio encoder in this artifact is the merged native checkpoint:
+`mn20_native_merged_aistmix_audioheavy100k175k175k_continue_from_balanced_20260426T143137Z/latest_model.pt`
+## Parameter Count
+| Component | Params |
+|---|---:|
+| Text encoder (`MongoDB/mdbr-leaf-ir`) | 22,861,056 |
+| Image encoder (`mobilenetv4_conv_medium.e180_r384_in12k`) | 8,434,512 |
+| Audio encoder (merged native `mn20_as`) | 19,886,566 |
+| Image projection head | 12,306,560 |
+| Audio projection head | 12,306,560 |
+| Text projection head | 11,323,520 |
+| **Total exact loaded params** | **87,118,774** |
+## Files
+| File | Purpose |
+|---|---|
+| `AIST-87M.safetensors` | Self-contained release artifact |
+| `aist_81m_raw_mn20_lora.yaml` | Training recipe for the source run |
+| `parameter_breakdown.json` | Exact parameter accounting |
+| `aist87m_memory_slice_release_report.md` | Human-memory slice report |
+| `aist87m_memory_slice_release_report.json` | Machine-readable evaluation summary |
+## Caveats
+- The model is optimized and reported for memory-relevant embedding surfaces,
+  not broad leaderboard coverage.
+- The single-audio path is smaller and simpler than the dual-audio tower, but
+  it does not dominate the dual-audio tower on paired diagnostic scores.
+- 1280d, 768d, and 512d human-memory slices are complete for the release checkpoint.

aist87m_memory_slice_release_report.json ADDED Viewed

	@@ -0,0 +1,1275 @@

+{
+  "tasks": [
+    "SprintDuplicateQuestions",
+    "STSBenchmark",
+    "Flickr30kT2IRetrieval",
+    "Flickr30kI2TRetrieval",
+    "CommonVoiceMini21T2ARetrieval",
+    "MACST2ARetrieval",
+    "UrbanSound8KT2ARetrieval",
+    "ClothoT2ARetrieval"
+  ],
+  "primary_metric_policy": {
+    "text": "main_score",
+    "image_text_retrieval": "ndcg_at_10",
+    "audio_text_retrieval": "ndcg_at_10"
+  },
+  "runs": [
+    {
+      "label": "AIST-87M 1280",
+      "dimension": 1280,
+      "results_dir": "/shared/augmem/triembed/results/aist81m_raw1280_mn20_merged_teacher_20260503T0125Z_memory_slice_default/dim1280/results/triembed__te-1280d/best_model",
+      "completed_tasks": 8,
+      "missing_tasks": [],
+      "overall_mean": 0.3491338344017094,
+      "family_means": {
+        "Audio recall": 0.1041809188034188,
+        "Image recall": 0.424995,
+        "Text continuity": 0.7631785
+      },
+      "rows": [
+        {
+          "label": "AIST-87M 1280",
+          "dimension": 1280,
+          "task": "SprintDuplicateQuestions",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.875145,
+          "metrics": {
+            "main_score": 0.875145
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 1280",
+          "dimension": 1280,
+          "task": "STSBenchmark",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.651212,
+          "metrics": {
+            "main_score": 0.651212,
+            "cosine_spearman": 0.651212,
+            "spearman": 0.651212
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 1280",
+          "dimension": 1280,
+          "task": "Flickr30kT2IRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.4685,
+          "metrics": {
+            "main_score": 0.4685,
+            "ndcg_at_10": 0.4685,
+            "recall_at_1": 0.2956,
+            "recall_at_10": 0.6718,
+            "mrr_at_10": 0.405197
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 1280",
+          "dimension": 1280,
+          "task": "Flickr30kI2TRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.38149,
+          "metrics": {
+            "main_score": 0.38149,
+            "ndcg_at_10": 0.38149,
+            "recall_at_1": 0.0816,
+            "recall_at_10": 0.4072,
+            "mrr_at_10": 0.533862
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 1280",
+          "dimension": 1280,
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.028403675213675213,
+          "metrics": {
+            "main_score": 0.03276290598290598,
+            "ndcg_at_10": 0.028403675213675213,
+            "recall_at_1": 0.005908376068376069,
+            "recall_at_10": 0.061962393162393166,
+            "mrr_at_10": 0.01842434188034188
+          },
+          "subsets": 117
+        },
+        {
+          "label": "AIST-87M 1280",
+          "dimension": 1280,
+          "task": "MACST2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.11037,
+          "metrics": {
+            "main_score": 0.13995,
+            "ndcg_at_10": 0.11037,
+            "recall_at_1": 0.03308,
+            "recall_at_10": 0.21374,
+            "mrr_at_10": 0.079078
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 1280",
+          "dimension": 1280,
+          "task": "UrbanSound8KT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.00851,
+          "metrics": {
+            "main_score": 0.00963,
+            "ndcg_at_10": 0.00851,
+            "recall_at_1": 0.00196,
+            "recall_at_10": 0.01847,
+            "mrr_at_10": 0.00556
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 1280",
+          "dimension": 1280,
+          "task": "ClothoT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.26944,
+          "metrics": {
+            "main_score": 0.3325,
+            "ndcg_at_10": 0.26944,
+            "recall_at_1": 0.1282,
+            "recall_at_10": 0.44315,
+            "mrr_at_10": 0.215861
+          },
+          "subsets": 1
+        }
+      ],
+      "source_result_dirs": [
+        "/shared/augmem/triembed/results/aist81m_raw1280_mn20_merged_teacher_20260503T0125Z_memory_slice_default/dim1280/results/triembed__te-1280d/best_model"
+      ]
+    },
+    {
+      "label": "AIST-87M 768",
+      "dimension": 768,
+      "results_dir": "/shared/augmem/triembed/results/aist81m_raw1280_mn20_merged_teacher_20260503T0125Z_memory_slice_default/dim768/results/triembed__te-768d/best_model",
+      "completed_tasks": 8,
+      "missing_tasks": [],
+      "overall_mean": 0.34871195512820513,
+      "family_means": {
+        "Audio recall": 0.10426891025641025,
+        "Image recall": 0.423815,
+        "Text continuity": 0.7624949999999999
+      },
+      "rows": [
+        {
+          "label": "AIST-87M 768",
+          "dimension": 768,
+          "task": "SprintDuplicateQuestions",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.874231,
+          "metrics": {
+            "main_score": 0.874231
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 768",
+          "dimension": 768,
+          "task": "STSBenchmark",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.650759,
+          "metrics": {
+            "main_score": 0.650759,
+            "cosine_spearman": 0.650759,
+            "spearman": 0.650759
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 768",
+          "dimension": 768,
+          "task": "Flickr30kT2IRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.46701,
+          "metrics": {
+            "main_score": 0.46701,
+            "ndcg_at_10": 0.46701,
+            "recall_at_1": 0.2922,
+            "recall_at_10": 0.6712,
+            "mrr_at_10": 0.403385
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 768",
+          "dimension": 768,
+          "task": "Flickr30kI2TRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.38062,
+          "metrics": {
+            "main_score": 0.38062,
+            "ndcg_at_10": 0.38062,
+            "recall_at_1": 0.0814,
+            "recall_at_10": 0.4058,
+            "mrr_at_10": 0.532687
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 768",
+          "dimension": 768,
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.028395641025641027,
+          "metrics": {
+            "main_score": 0.03299991452991453,
+            "ndcg_at_10": 0.028395641025641027,
+            "recall_at_1": 0.005907350427350427,
+            "recall_at_10": 0.062035897435897436,
+            "mrr_at_10": 0.01839460683760684
+          },
+          "subsets": 117
+        },
+        {
+          "label": "AIST-87M 768",
+          "dimension": 768,
+          "task": "MACST2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.11149,
+          "metrics": {
+            "main_score": 0.14249,
+            "ndcg_at_10": 0.11149,
+            "recall_at_1": 0.03308,
+            "recall_at_10": 0.21628,
+            "mrr_at_10": 0.079723
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 768",
+          "dimension": 768,
+          "task": "UrbanSound8KT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.00851,
+          "metrics": {
+            "main_score": 0.00963,
+            "ndcg_at_10": 0.00851,
+            "recall_at_1": 0.00196,
+            "recall_at_10": 0.01847,
+            "mrr_at_10": 0.005562
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 768",
+          "dimension": 768,
+          "task": "ClothoT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.26868,
+          "metrics": {
+            "main_score": 0.33178,
+            "ndcg_at_10": 0.26868,
+            "recall_at_1": 0.12695,
+            "recall_at_10": 0.44208,
+            "mrr_at_10": 0.21516
+          },
+          "subsets": 1
+        }
+      ],
+      "source_result_dirs": [
+        "/shared/augmem/triembed/results/aist81m_raw1280_mn20_merged_teacher_20260503T0125Z_memory_slice_default/dim768/results/triembed__te-768d/best_model",
+        "/shared/augmem/triembed/results/aist81m_raw1280_mn20_merged_teacher_20260503T0125Z_memory_slice_dim768_fill/dim768/results/triembed__te-768d/best_model"
+      ]
+    },
+    {
+      "label": "AIST-87M 512",
+      "dimension": 512,
+      "results_dir": "/shared/augmem/triembed/results/aist81m_raw1280_mn20_merged_teacher_20260503T0125Z_memory_slice_dim512/dim512/results/triembed__te-512d/best_model",
+      "completed_tasks": 8,
+      "missing_tasks": [],
+      "overall_mean": 0.3488224732905983,
+      "family_means": {
+        "Audio recall": 0.10438869658119658,
+        "Image recall": 0.42417499999999997,
+        "Text continuity": 0.7623375
+      },
+      "rows": [
+        {
+          "label": "AIST-87M 512",
+          "dimension": 512,
+          "task": "SprintDuplicateQuestions",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.873508,
+          "metrics": {
+            "main_score": 0.873508
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 512",
+          "dimension": 512,
+          "task": "STSBenchmark",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.651167,
+          "metrics": {
+            "main_score": 0.651167,
+            "cosine_spearman": 0.651167,
+            "spearman": 0.651167
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 512",
+          "dimension": 512,
+          "task": "Flickr30kT2IRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.4676,
+          "metrics": {
+            "main_score": 0.4676,
+            "ndcg_at_10": 0.4676,
+            "recall_at_1": 0.2954,
+            "recall_at_10": 0.6702,
+            "mrr_at_10": 0.404515
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 512",
+          "dimension": 512,
+          "task": "Flickr30kI2TRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.38075,
+          "metrics": {
+            "main_score": 0.38075,
+            "ndcg_at_10": 0.38075,
+            "recall_at_1": 0.0824,
+            "recall_at_10": 0.4052,
+            "mrr_at_10": 0.535146
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 512",
+          "dimension": 512,
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.028264786324786326,
+          "metrics": {
+            "main_score": 0.03229504273504274,
+            "ndcg_at_10": 0.028264786324786326,
+            "recall_at_1": 0.006467948717948718,
+            "recall_at_10": 0.060837521367521366,
+            "mrr_at_10": 0.018573598290598292
+          },
+          "subsets": 117
+        },
+        {
+          "label": "AIST-87M 512",
+          "dimension": 512,
+          "task": "MACST2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.11287,
+          "metrics": {
+            "main_score": 0.13486,
+            "ndcg_at_10": 0.11287,
+            "recall_at_1": 0.03308,
+            "recall_at_10": 0.22137,
+            "mrr_at_10": 0.080181
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 512",
+          "dimension": 512,
+          "task": "UrbanSound8KT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.0085,
+          "metrics": {
+            "main_score": 0.00923,
+            "ndcg_at_10": 0.0085,
+            "recall_at_1": 0.00196,
+            "recall_at_10": 0.01847,
+            "mrr_at_10": 0.005544
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-87M 512",
+          "dimension": 512,
+          "task": "ClothoT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.26792,
+          "metrics": {
+            "main_score": 0.33107,
+            "ndcg_at_10": 0.26792,
+            "recall_at_1": 0.1248,
+            "recall_at_10": 0.44261,
+            "mrr_at_10": 0.213985
+          },
+          "subsets": 1
+        }
+      ],
+      "source_result_dirs": [
+        "/shared/augmem/triembed/results/aist81m_raw1280_mn20_merged_teacher_20260503T0125Z_memory_slice_dim512/dim512/results/triembed__te-512d/best_model"
+      ]
+    },
+    {
+      "label": "AIST-95M 1280 Flickr",
+      "dimension": 1280,
+      "results_dir": "/shared/augmem/triembed/results/aist95m_1280_mieb_flickr_20260502T0217Z/dim1280/results/triembed__te-1280d/best_model",
+      "completed_tasks": 2,
+      "missing_tasks": [
+        "ClothoT2ARetrieval",
+        "CommonVoiceMini21T2ARetrieval",
+        "MACST2ARetrieval",
+        "STSBenchmark",
+        "SprintDuplicateQuestions",
+        "UrbanSound8KT2ARetrieval"
+      ],
+      "overall_mean": 0.485,
+      "family_means": {
+        "Image recall": 0.485
+      },
+      "rows": [
+        {
+          "label": "AIST-95M 1280 Flickr",
+          "dimension": 1280,
+          "task": "Flickr30kT2IRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.50216,
+          "metrics": {
+            "main_score": 0.50216,
+            "ndcg_at_10": 0.50216,
+            "recall_at_1": 0.3254,
+            "recall_at_10": 0.7004,
+            "mrr_at_10": 0.439975
+          },
+          "subsets": 1
+        },
+        {
+          "label": "AIST-95M 1280 Flickr",
+          "dimension": 1280,
+          "task": "Flickr30kI2TRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.46784,
+          "metrics": {
+            "main_score": 0.46784,
+            "ndcg_at_10": 0.46784,
+            "recall_at_1": 0.0958,
+            "recall_at_10": 0.5034,
+            "mrr_at_10": 0.598869
+          },
+          "subsets": 1
+        }
+      ],
+      "source_result_dirs": [
+        "/shared/augmem/triembed/results/aist95m_1280_mieb_flickr_20260502T0217Z/dim1280/results/triembed__te-1280d/best_model"
+      ]
+    },
+    {
+      "label": "ES-AIST-81M 768",
+      "dimension": 768,
+      "results_dir": "/shared/augmem/triembed/results/es_aist_memory_slice_default_20260501T1835Z/dim768/results/triembed__te-768d/best_model",
+      "completed_tasks": 8,
+      "missing_tasks": [],
+      "overall_mean": 0.30764677777777777,
+      "family_means": {
+        "Audio recall": 0.06462555555555556,
+        "Image recall": 0.271195,
+        "Text continuity": 0.830141
+      },
+      "rows": [
+        {
+          "label": "ES-AIST-81M 768",
+          "dimension": 768,
+          "task": "SprintDuplicateQuestions",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.916128,
+          "metrics": {
+            "main_score": 0.916128
+          },
+          "subsets": 1
+        },
+        {
+          "label": "ES-AIST-81M 768",
+          "dimension": 768,
+          "task": "STSBenchmark",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.744154,
+          "metrics": {
+            "main_score": 0.744154,
+            "cosine_spearman": 0.744154,
+            "spearman": 0.744154
+          },
+          "subsets": 1
+        },
+        {
+          "label": "ES-AIST-81M 768",
+          "dimension": 768,
+          "task": "Flickr30kT2IRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.34676,
+          "metrics": {
+            "main_score": 0.34676,
+            "ndcg_at_10": 0.34676,
+            "recall_at_1": 0.1764,
+            "recall_at_10": 0.5528,
+            "mrr_at_10": 0.282987
+          },
+          "subsets": 1
+        },
+        {
+          "label": "ES-AIST-81M 768",
+          "dimension": 768,
+          "task": "Flickr30kI2TRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.19563,
+          "metrics": {
+            "main_score": 0.19563,
+            "ndcg_at_10": 0.19563,
+            "recall_at_1": 0.037,
+            "recall_at_10": 0.2208,
+            "mrr_at_10": 0.295532
+          },
+          "subsets": 1
+        },
+        {
+          "label": "ES-AIST-81M 768",
+          "dimension": 768,
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.024182222222222223,
+          "metrics": {
+            "main_score": 0.02774760683760684,
+            "ndcg_at_10": 0.024182222222222223,
+            "recall_at_1": 0.005472478632478632,
+            "recall_at_10": 0.052400170940170944,
+            "mrr_at_10": 0.01579925641025641
+          },
+          "subsets": 117
+        },
+        {
+          "label": "ES-AIST-81M 768",
+          "dimension": 768,
+          "task": "MACST2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.07729,
+          "metrics": {
+            "main_score": 0.08906,
+            "ndcg_at_10": 0.07729,
+            "recall_at_1": 0.0229,
+            "recall_at_10": 0.15013,
+            "mrr_at_10": 0.055219
+          },
+          "subsets": 1
+        },
+        {
+          "label": "ES-AIST-81M 768",
+          "dimension": 768,
+          "task": "UrbanSound8KT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.007,
+          "metrics": {
+            "main_score": 0.00747,
+            "ndcg_at_10": 0.007,
+            "recall_at_1": 0.00098,
+            "recall_at_10": 0.01631,
+            "mrr_at_10": 0.004257
+          },
+          "subsets": 1
+        },
+        {
+          "label": "ES-AIST-81M 768",
+          "dimension": 768,
+          "task": "ClothoT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.15003,
+          "metrics": {
+            "main_score": 0.17601,
+            "ndcg_at_10": 0.15003,
+            "recall_at_1": 0.05121,
+            "recall_at_10": 0.28612,
+            "mrr_at_10": 0.108968
+          },
+          "subsets": 1
+        }
+      ],
+      "source_result_dirs": [
+        "/shared/augmem/triembed/results/es_aist_memory_slice_default_20260501T1835Z/dim768/results/triembed__te-768d/best_model"
+      ]
+    },
+    {
+      "label": "Native mn20 audio 768",
+      "dimension": 768,
+      "results_dir": "/shared/augmem/triembed/results/es_aist_memory_audio_native_default_20260501T1835Z/dim768/results/triembed__native-efficientat-768d/latest_model",
+      "completed_tasks": 4,
+      "missing_tasks": [
+        "Flickr30kI2TRetrieval",
+        "Flickr30kT2IRetrieval",
+        "STSBenchmark",
+        "SprintDuplicateQuestions"
+      ],
+      "overall_mean": 0.11513626068376069,
+      "family_means": {
+        "Audio recall": 0.11513626068376069
+      },
+      "rows": [
+        {
+          "label": "Native mn20 audio 768",
+          "dimension": 768,
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.035825042735042736,
+          "metrics": {
+            "main_score": 0.04166820512820513,
+            "ndcg_at_10": 0.035825042735042736,
+            "recall_at_1": 0.009125726495726495,
+            "recall_at_10": 0.07585017094017094,
+            "mrr_at_10": 0.023907692307692307
+          },
+          "subsets": 117
+        },
+        {
+          "label": "Native mn20 audio 768",
+          "dimension": 768,
+          "task": "MACST2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.12746,
+          "metrics": {
+            "main_score": 0.13995,
+            "ndcg_at_10": 0.12746,
+            "recall_at_1": 0.05852,
+            "recall_at_10": 0.22392,
+            "mrr_at_10": 0.098715
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Native mn20 audio 768",
+          "dimension": 768,
+          "task": "UrbanSound8KT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.00849,
+          "metrics": {
+            "main_score": 0.00923,
+            "ndcg_at_10": 0.00849,
+            "recall_at_1": 0.00196,
+            "recall_at_10": 0.01866,
+            "mrr_at_10": 0.005487
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Native mn20 audio 768",
+          "dimension": 768,
+          "task": "ClothoT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.28877,
+          "metrics": {
+            "main_score": 0.3581,
+            "ndcg_at_10": 0.28877,
+            "recall_at_1": 0.14414,
+            "recall_at_10": 0.4641,
+            "mrr_at_10": 0.234475
+          },
+          "subsets": 1
+        }
+      ],
+      "source_result_dirs": [
+        "/shared/augmem/triembed/results/es_aist_memory_audio_native_default_20260501T1835Z/dim768/results/triembed__native-efficientat-768d/latest_model"
+      ]
+    },
+    {
+      "label": "Dual-audio tower 1280",
+      "dimension": 1280,
+      "results_dir": "/shared/augmem/triembed/results/aist86m_full_mteb_mieb_maeb_1280_768_512_20260502T070609Z/dim1280/results/triembed__te-1280d/TE-86M-dual-audio-best_model",
+      "completed_tasks": 8,
+      "missing_tasks": [],
+      "overall_mean": 0.3973782852564103,
+      "family_means": {
+        "Audio recall": 0.11287532051282051,
+        "Image recall": 0.485,
+        "Text continuity": 0.8787625
+      },
+      "rows": [
+        {
+          "label": "Dual-audio tower 1280",
+          "dimension": 1280,
+          "task": "SprintDuplicateQuestions",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.953368,
+          "metrics": {
+            "main_score": 0.953368
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 1280",
+          "dimension": 1280,
+          "task": "STSBenchmark",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.804157,
+          "metrics": {
+            "main_score": 0.804157,
+            "cosine_spearman": 0.804157,
+            "spearman": 0.804154
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 1280",
+          "dimension": 1280,
+          "task": "Flickr30kT2IRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.50216,
+          "metrics": {
+            "main_score": 0.50216,
+            "ndcg_at_10": 0.50216,
+            "recall_at_1": 0.3254,
+            "recall_at_10": 0.7004,
+            "mrr_at_10": 0.439975
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 1280",
+          "dimension": 1280,
+          "task": "Flickr30kI2TRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.46784,
+          "metrics": {
+            "main_score": 0.46784,
+            "ndcg_at_10": 0.46784,
+            "recall_at_1": 0.0958,
+            "recall_at_10": 0.5034,
+            "mrr_at_10": 0.598869
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 1280",
+          "dimension": 1280,
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.03849128205128205,
+          "metrics": {
+            "main_score": 0.04426282051282051,
+            "ndcg_at_10": 0.03849128205128205,
+            "recall_at_1": 0.00971991452991453,
+            "recall_at_10": 0.08076905982905982,
+            "mrr_at_10": 0.02587371794871795
+          },
+          "subsets": 117
+        },
+        {
+          "label": "Dual-audio tower 1280",
+          "dimension": 1280,
+          "task": "MACST2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.10964,
+          "metrics": {
+            "main_score": 0.15522,
+            "ndcg_at_10": 0.10964,
+            "recall_at_1": 0.04326,
+            "recall_at_10": 0.19338,
+            "mrr_at_10": 0.083683
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 1280",
+          "dimension": 1280,
+          "task": "UrbanSound8KT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.00823,
+          "metrics": {
+            "main_score": 0.00904,
+            "ndcg_at_10": 0.00823,
+            "recall_at_1": 0.00177,
+            "recall_at_10": 0.01807,
+            "mrr_at_10": 0.00531
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 1280",
+          "dimension": 1280,
+          "task": "ClothoT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.29514,
+          "metrics": {
+            "main_score": 0.36043,
+            "ndcg_at_10": 0.29514,
+            "recall_at_1": 0.14861,
+            "recall_at_10": 0.47395,
+            "mrr_at_10": 0.239903
+          },
+          "subsets": 1
+        }
+      ],
+      "source_result_dirs": [
+        "/shared/augmem/triembed/results/aist86m_full_mteb_mieb_maeb_1280_768_512_20260502T070609Z/dim1280/results/triembed__te-1280d/TE-86M-dual-audio-best_model"
+      ]
+    },
+    {
+      "label": "Dual-audio tower 768",
+      "dimension": 768,
+      "results_dir": "/shared/augmem/triembed/results/aist86m_full_mteb_mieb_maeb_1280_768_512_20260502T070609Z/dim768/results/triembed__te-768d/TE-86M-dual-audio-best_model",
+      "completed_tasks": 6,
+      "missing_tasks": [
+        "MACST2ARetrieval",
+        "UrbanSound8KT2ARetrieval"
+      ],
+      "overall_mean": 0.5098147193732193,
+      "family_means": {
+        "Audio recall": 0.16678465811965812,
+        "Image recall": 0.48403999999999997,
+        "Text continuity": 0.8786195
+      },
+      "rows": [
+        {
+          "label": "Dual-audio tower 768",
+          "dimension": 768,
+          "task": "SprintDuplicateQuestions",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.953072,
+          "metrics": {
+            "main_score": 0.953072
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 768",
+          "dimension": 768,
+          "task": "STSBenchmark",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.804167,
+          "metrics": {
+            "main_score": 0.804167,
+            "cosine_spearman": 0.804167,
+            "spearman": 0.804167
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 768",
+          "dimension": 768,
+          "task": "Flickr30kT2IRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.50179,
+          "metrics": {
+            "main_score": 0.50179,
+            "ndcg_at_10": 0.50179,
+            "recall_at_1": 0.3254,
+            "recall_at_10": 0.698,
+            "mrr_at_10": 0.440147
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 768",
+          "dimension": 768,
+          "task": "Flickr30kI2TRetrieval",
+          "family": "Image recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.46629,
+          "metrics": {
+            "main_score": 0.46629,
+            "ndcg_at_10": 0.46629,
+            "recall_at_1": 0.0956,
+            "recall_at_10": 0.5022,
+            "mrr_at_10": 0.597365
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 768",
+          "dimension": 768,
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.03849931623931624,
+          "metrics": {
+            "main_score": 0.04466316239316239,
+            "ndcg_at_10": 0.03849931623931624,
+            "recall_at_1": 0.009814871794871794,
+            "recall_at_10": 0.08058384615384616,
+            "mrr_at_10": 0.025928871794871796
+          },
+          "subsets": 117
+        },
+        {
+          "label": "Dual-audio tower 768",
+          "dimension": 768,
+          "task": "ClothoT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.29507,
+          "metrics": {
+            "main_score": 0.3615,
+            "ndcg_at_10": 0.29507,
+            "recall_at_1": 0.14861,
+            "recall_at_10": 0.47359,
+            "mrr_at_10": 0.239883
+          },
+          "subsets": 1
+        }
+      ],
+      "source_result_dirs": [
+        "/shared/augmem/triembed/results/aist86m_full_mteb_mieb_maeb_1280_768_512_20260502T070609Z/dim768/results/triembed__te-768d/TE-86M-dual-audio-best_model"
+      ]
+    },
+    {
+      "label": "Dual-audio tower 512",
+      "dimension": 512,
+      "results_dir": "/shared/augmem/triembed/results/aist86m_full_mteb_mieb_maeb_1280_768_512_20260502T070609Z/dim512/results/triembed__te-512d/TE-86M-dual-audio-best_model",
+      "completed_tasks": 4,
+      "missing_tasks": [
+        "Flickr30kI2TRetrieval",
+        "Flickr30kT2IRetrieval",
+        "MACST2ARetrieval",
+        "UrbanSound8KT2ARetrieval"
+      ],
+      "overall_mean": 0.5228179594017094,
+      "family_means": {
+        "Audio recall": 0.16697341880341882,
+        "Text continuity": 0.8786625
+      },
+      "rows": [
+        {
+          "label": "Dual-audio tower 512",
+          "dimension": 512,
+          "task": "SprintDuplicateQuestions",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.952893,
+          "metrics": {
+            "main_score": 0.952893
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 512",
+          "dimension": 512,
+          "task": "STSBenchmark",
+          "family": "Text continuity",
+          "primary_metric": "main_score",
+          "primary": 0.804432,
+          "metrics": {
+            "main_score": 0.804432,
+            "cosine_spearman": 0.804432,
+            "spearman": 0.804432
+          },
+          "subsets": 1
+        },
+        {
+          "label": "Dual-audio tower 512",
+          "dimension": 512,
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.03858683760683761,
+          "metrics": {
+            "main_score": 0.04408854700854701,
+            "ndcg_at_10": 0.03858683760683761,
+            "recall_at_1": 0.00959076923076923,
+            "recall_at_10": 0.08129623931623932,
+            "mrr_at_10": 0.025843299145299144
+          },
+          "subsets": 117
+        },
+        {
+          "label": "Dual-audio tower 512",
+          "dimension": 512,
+          "task": "ClothoT2ARetrieval",
+          "family": "Audio recall",
+          "primary_metric": "ndcg_at_10",
+          "primary": 0.29536,
+          "metrics": {
+            "main_score": 0.35882,
+            "ndcg_at_10": 0.29536,
+            "recall_at_1": 0.1513,
+            "recall_at_10": 0.47162,
+            "mrr_at_10": 0.240905
+          },
+          "subsets": 1
+        }
+      ],
+      "source_result_dirs": [
+        "/shared/augmem/triembed/results/aist86m_full_mteb_mieb_maeb_1280_768_512_20260502T070609Z/dim512/results/triembed__te-512d/TE-86M-dual-audio-best_model"
+      ]
+    }
+  ],
+  "comparisons": [
+    {
+      "baseline": "ES-AIST-81M 768",
+      "target": "AIST-87M 768",
+      "paired_tasks": 8,
+      "mean_absolute_delta": 0.041065177350427334,
+      "rows": [
+        {
+          "task": "SprintDuplicateQuestions",
+          "dimension": 768,
+          "family": "Text continuity",
+          "baseline": "ES-AIST-81M 768",
+          "baseline_primary": 0.916128,
+          "target": "AIST-87M 768",
+          "target_primary": 0.874231,
+          "absolute_delta": -0.04189700000000007,
+          "relative_delta_pct": -4.573269237486473
+        },
+        {
+          "task": "STSBenchmark",
+          "dimension": 768,
+          "family": "Text continuity",
+          "baseline": "ES-AIST-81M 768",
+          "baseline_primary": 0.744154,
+          "target": "AIST-87M 768",
+          "target_primary": 0.650759,
+          "absolute_delta": -0.093395,
+          "relative_delta_pct": -12.550493580629817
+        },
+        {
+          "task": "Flickr30kT2IRetrieval",
+          "dimension": 768,
+          "family": "Image recall",
+          "baseline": "ES-AIST-81M 768",
+          "baseline_primary": 0.34676,
+          "target": "AIST-87M 768",
+          "target_primary": 0.46701,
+          "absolute_delta": 0.12024999999999997,
+          "relative_delta_pct": 34.67816357134617
+        },
+        {
+          "task": "Flickr30kI2TRetrieval",
+          "dimension": 768,
+          "family": "Image recall",
+          "baseline": "ES-AIST-81M 768",
+          "baseline_primary": 0.19563,
+          "target": "AIST-87M 768",
+          "target_primary": 0.38062,
+          "absolute_delta": 0.18499000000000002,
+          "relative_delta_pct": 94.56116137606708
+        },
+        {
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "dimension": 768,
+          "family": "Audio recall",
+          "baseline": "ES-AIST-81M 768",
+          "baseline_primary": 0.024182222222222223,
+          "target": "AIST-87M 768",
+          "target_primary": 0.028395641025641027,
+          "absolute_delta": 0.004213418803418804,
+          "relative_delta_pct": 17.423621223474193
+        },
+        {
+          "task": "MACST2ARetrieval",
+          "dimension": 768,
+          "family": "Audio recall",
+          "baseline": "ES-AIST-81M 768",
+          "baseline_primary": 0.07729,
+          "target": "AIST-87M 768",
+          "target_primary": 0.11149,
+          "absolute_delta": 0.03420000000000001,
+          "relative_delta_pct": 44.24893259153838
+        },
+        {
+          "task": "UrbanSound8KT2ARetrieval",
+          "dimension": 768,
+          "family": "Audio recall",
+          "baseline": "ES-AIST-81M 768",
+          "baseline_primary": 0.007,
+          "target": "AIST-87M 768",
+          "target_primary": 0.00851,
+          "absolute_delta": 0.00151,
+          "relative_delta_pct": 21.571428571428573
+        },
+        {
+          "task": "ClothoT2ARetrieval",
+          "dimension": 768,
+          "family": "Audio recall",
+          "baseline": "ES-AIST-81M 768",
+          "baseline_primary": 0.15003,
+          "target": "AIST-87M 768",
+          "target_primary": 0.26868,
+          "absolute_delta": 0.11864999999999998,
+          "relative_delta_pct": 79.08418316336731
+        }
+      ]
+    },
+    {
+      "baseline": "Native mn20 audio 768",
+      "target": "AIST-87M 768",
+      "paired_tasks": 4,
+      "mean_absolute_delta": -0.010867350427350436,
+      "rows": [
+        {
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "dimension": 768,
+          "family": "Audio recall",
+          "baseline": "Native mn20 audio 768",
+          "baseline_primary": 0.035825042735042736,
+          "target": "AIST-87M 768",
+          "target_primary": 0.028395641025641027,
+          "absolute_delta": -0.00742940170940171,
+          "relative_delta_pct": -20.73801213399403
+        },
+        {
+          "task": "MACST2ARetrieval",
+          "dimension": 768,
+          "family": "Audio recall",
+          "baseline": "Native mn20 audio 768",
+          "baseline_primary": 0.12746,
+          "target": "AIST-87M 768",
+          "target_primary": 0.11149,
+          "absolute_delta": -0.015969999999999984,
+          "relative_delta_pct": -12.529420994821894
+        },
+        {
+          "task": "UrbanSound8KT2ARetrieval",
+          "dimension": 768,
+          "family": "Audio recall",
+          "baseline": "Native mn20 audio 768",
+          "baseline_primary": 0.00849,
+          "target": "AIST-87M 768",
+          "target_primary": 0.00851,
+          "absolute_delta": 2.000000000000092e-05,
+          "relative_delta_pct": 0.2355712603062535
+        },
+        {
+          "task": "ClothoT2ARetrieval",
+          "dimension": 768,
+          "family": "Audio recall",
+          "baseline": "Native mn20 audio 768",
+          "baseline_primary": 0.28877,
+          "target": "AIST-87M 768",
+          "target_primary": 0.26868,
+          "absolute_delta": -0.020090000000000052,
+          "relative_delta_pct": -6.9570938809433285
+        }
+      ]
+    },
+    {
+      "baseline": "Dual-audio tower 768",
+      "target": "AIST-87M 768",
+      "paired_tasks": 6,
+      "mean_absolute_delta": -0.06486544586894587,
+      "rows": [
+        {
+          "task": "SprintDuplicateQuestions",
+          "dimension": 768,
+          "family": "Text continuity",
+          "baseline": "Dual-audio tower 768",
+          "baseline_primary": 0.953072,
+          "target": "AIST-87M 768",
+          "target_primary": 0.874231,
+          "absolute_delta": -0.07884100000000005,
+          "relative_delta_pct": -8.272302617220948
+        },
+        {
+          "task": "STSBenchmark",
+          "dimension": 768,
+          "family": "Text continuity",
+          "baseline": "Dual-audio tower 768",
+          "baseline_primary": 0.804167,
+          "target": "AIST-87M 768",
+          "target_primary": 0.650759,
+          "absolute_delta": -0.153408,
+          "relative_delta_pct": -19.076634579633335
+        },
+        {
+          "task": "Flickr30kT2IRetrieval",
+          "dimension": 768,
+          "family": "Image recall",
+          "baseline": "Dual-audio tower 768",
+          "baseline_primary": 0.50179,
+          "target": "AIST-87M 768",
+          "target_primary": 0.46701,
+          "absolute_delta": -0.03477999999999998,
+          "relative_delta_pct": -6.931186352856769
+        },
+        {
+          "task": "Flickr30kI2TRetrieval",
+          "dimension": 768,
+          "family": "Image recall",
+          "baseline": "Dual-audio tower 768",
+          "baseline_primary": 0.46629,
+          "target": "AIST-87M 768",
+          "target_primary": 0.38062,
+          "absolute_delta": -0.08566999999999997,
+          "relative_delta_pct": -18.372686525552762
+        },
+        {
+          "task": "CommonVoiceMini21T2ARetrieval",
+          "dimension": 768,
+          "family": "Audio recall",
+          "baseline": "Dual-audio tower 768",
+          "baseline_primary": 0.03849931623931624,
+          "target": "AIST-87M 768",
+          "target_primary": 0.028395641025641027,
+          "absolute_delta": -0.010103675213675212,
+          "relative_delta_pct": -26.243778333281533
+        },
+        {
+          "task": "ClothoT2ARetrieval",
+          "dimension": 768,
+          "family": "Audio recall",
+          "baseline": "Dual-audio tower 768",
+          "baseline_primary": 0.29507,
+          "target": "AIST-87M 768",
+          "target_primary": 0.26868,
+          "absolute_delta": -0.026390000000000025,
+          "relative_delta_pct": -8.943640492086633
+        }
+      ]
+    }
+  ]
+}

aist87m_memory_slice_release_report.md ADDED Viewed

	@@ -0,0 +1,85 @@

+# AIST-87M Human-Memory Evaluation Slice
+Primary metrics are `main_score` for text continuity tasks and `NDCG@10` for image/audio retrieval tasks.
+| Run | Dim | Tasks | Text continuity | Image recall | Audio recall | Overall | Missing |
+|---|---:|---:|---:|---:|---:|---:|---|
+| AIST-87M 1280 | 1280 | 8 | 0.763 | 0.425 | 0.104 | 0.349 | none |
+| AIST-87M 768 | 768 | 8 | 0.762 | 0.424 | 0.104 | 0.349 | none |
+| AIST-87M 512 | 512 | 8 | 0.762 | 0.424 | 0.104 | 0.349 | none |
+| AIST-95M 1280 Flickr | 1280 | 2 | - | 0.485 | - | 0.485 | ClothoT2ARetrieval, CommonVoiceMini21T2ARetrieval, MACST2ARetrieval, STSBenchmark, SprintDuplicateQuestions, UrbanSound8KT2ARetrieval |
+| ES-AIST-81M 768 | 768 | 8 | 0.830 | 0.271 | 0.065 | 0.308 | none |
+| Native mn20 audio 768 | 768 | 4 | - | - | 0.115 | 0.115 | Flickr30kI2TRetrieval, Flickr30kT2IRetrieval, STSBenchmark, SprintDuplicateQuestions |
+| Dual-audio tower 1280 | 1280 | 8 | 0.879 | 0.485 | 0.113 | 0.397 | none |
+| Dual-audio tower 768 | 768 | 6 | 0.879 | 0.484 | 0.167 | 0.510 | MACST2ARetrieval, UrbanSound8KT2ARetrieval |
+| Dual-audio tower 512 | 512 | 4 | 0.879 | - | 0.167 | 0.523 | Flickr30kI2TRetrieval, Flickr30kT2IRetrieval, MACST2ARetrieval, UrbanSound8KT2ARetrieval |
+## AIST-87M Per-Task Scores
+| Dim | Task | Family | Metric | Score | R@1 | R@10 | MRR@10 |
+|---:|---|---|---|---:|---:|---:|---:|
+| 1280 | SprintDuplicateQuestions | Text continuity | main_score | 0.875 | - | - | - |
+| 1280 | STSBenchmark | Text continuity | main_score | 0.651 | - | - | - |
+| 1280 | Flickr30kT2IRetrieval | Image recall | ndcg_at_10 | 0.469 | 0.296 | 0.672 | 0.405 |
+| 1280 | Flickr30kI2TRetrieval | Image recall | ndcg_at_10 | 0.381 | 0.082 | 0.407 | 0.534 |
+| 1280 | CommonVoiceMini21T2ARetrieval | Audio recall | ndcg_at_10 | 0.028 | 0.006 | 0.062 | 0.018 |
+| 1280 | MACST2ARetrieval | Audio recall | ndcg_at_10 | 0.110 | 0.033 | 0.214 | 0.079 |
+| 1280 | UrbanSound8KT2ARetrieval | Audio recall | ndcg_at_10 | 0.009 | 0.002 | 0.018 | 0.006 |
+| 1280 | ClothoT2ARetrieval | Audio recall | ndcg_at_10 | 0.269 | 0.128 | 0.443 | 0.216 |
+| 768 | SprintDuplicateQuestions | Text continuity | main_score | 0.874 | - | - | - |
+| 768 | STSBenchmark | Text continuity | main_score | 0.651 | - | - | - |
+| 768 | Flickr30kT2IRetrieval | Image recall | ndcg_at_10 | 0.467 | 0.292 | 0.671 | 0.403 |
+| 768 | Flickr30kI2TRetrieval | Image recall | ndcg_at_10 | 0.381 | 0.081 | 0.406 | 0.533 |
+| 768 | CommonVoiceMini21T2ARetrieval | Audio recall | ndcg_at_10 | 0.028 | 0.006 | 0.062 | 0.018 |
+| 768 | MACST2ARetrieval | Audio recall | ndcg_at_10 | 0.111 | 0.033 | 0.216 | 0.080 |
+| 768 | UrbanSound8KT2ARetrieval | Audio recall | ndcg_at_10 | 0.009 | 0.002 | 0.018 | 0.006 |
+| 768 | ClothoT2ARetrieval | Audio recall | ndcg_at_10 | 0.269 | 0.127 | 0.442 | 0.215 |
+| 512 | SprintDuplicateQuestions | Text continuity | main_score | 0.874 | - | - | - |
+| 512 | STSBenchmark | Text continuity | main_score | 0.651 | - | - | - |
+| 512 | Flickr30kT2IRetrieval | Image recall | ndcg_at_10 | 0.468 | 0.295 | 0.670 | 0.405 |
+| 512 | Flickr30kI2TRetrieval | Image recall | ndcg_at_10 | 0.381 | 0.082 | 0.405 | 0.535 |
+| 512 | CommonVoiceMini21T2ARetrieval | Audio recall | ndcg_at_10 | 0.028 | 0.006 | 0.061 | 0.019 |
+| 512 | MACST2ARetrieval | Audio recall | ndcg_at_10 | 0.113 | 0.033 | 0.221 | 0.080 |
+| 512 | UrbanSound8KT2ARetrieval | Audio recall | ndcg_at_10 | 0.009 | 0.002 | 0.018 | 0.006 |
+| 512 | ClothoT2ARetrieval | Audio recall | ndcg_at_10 | 0.268 | 0.125 | 0.443 | 0.214 |
+## Paired Comparisons
+### AIST-87M 768 vs ES-AIST-81M 768
+Mean absolute delta over 8 paired tasks: 0.041.
+| Dim | Task | Baseline | Target | Absolute delta | Relative delta |
+|---:|---|---:|---:|---:|---:|
+| 768 | SprintDuplicateQuestions | 0.916 | 0.874 | -0.042 | -4.6% |
+| 768 | STSBenchmark | 0.744 | 0.651 | -0.093 | -12.6% |
+| 768 | Flickr30kT2IRetrieval | 0.347 | 0.467 | 0.120 | 34.7% |
+| 768 | Flickr30kI2TRetrieval | 0.196 | 0.381 | 0.185 | 94.6% |
+| 768 | CommonVoiceMini21T2ARetrieval | 0.024 | 0.028 | 0.004 | 17.4% |
+| 768 | MACST2ARetrieval | 0.077 | 0.111 | 0.034 | 44.2% |
+| 768 | UrbanSound8KT2ARetrieval | 0.007 | 0.009 | 0.002 | 21.6% |
+| 768 | ClothoT2ARetrieval | 0.150 | 0.269 | 0.119 | 79.1% |
+### AIST-87M 768 vs Native mn20 audio 768
+Mean absolute delta over 4 paired tasks: -0.011.
+| Dim | Task | Baseline | Target | Absolute delta | Relative delta |
+|---:|---|---:|---:|---:|---:|
+| 768 | CommonVoiceMini21T2ARetrieval | 0.036 | 0.028 | -0.007 | -20.7% |
+| 768 | MACST2ARetrieval | 0.127 | 0.111 | -0.016 | -12.5% |
+| 768 | UrbanSound8KT2ARetrieval | 0.008 | 0.009 | 0.000 | 0.2% |
+| 768 | ClothoT2ARetrieval | 0.289 | 0.269 | -0.020 | -7.0% |
+### AIST-87M 768 vs Dual-audio tower 768
+Mean absolute delta over 6 paired tasks: -0.065.
+| Dim | Task | Baseline | Target | Absolute delta | Relative delta |
+|---:|---|---:|---:|---:|---:|
+| 768 | SprintDuplicateQuestions | 0.953 | 0.874 | -0.079 | -8.3% |
+| 768 | STSBenchmark | 0.804 | 0.651 | -0.153 | -19.1% |
+| 768 | Flickr30kT2IRetrieval | 0.502 | 0.467 | -0.035 | -6.9% |
+| 768 | Flickr30kI2TRetrieval | 0.466 | 0.381 | -0.086 | -18.4% |
+| 768 | CommonVoiceMini21T2ARetrieval | 0.038 | 0.028 | -0.010 | -26.2% |
+| 768 | ClothoT2ARetrieval | 0.295 | 0.269 | -0.026 | -8.9% |

aist_81m_raw_mn20_lora.yaml ADDED Viewed

	@@ -0,0 +1,55 @@

+# Raw AIST-81M baseline.
+#
+# Generic trimodal InfoNCE teacher recipe using the single native mn20_as LoRA
+# audio path already used by ES-AIST-81M. This intentionally has no ES/entity
+# signal layout or entity-specific corpus/loss. It is the full-data baseline;
+# subset cache aliases are only for smoke tests.
+dataset_dir: datasets
+dataset_name: wordnet_2024_openai_validaudio
+cache_dir: cache
+encoder_name: mobilenetv4_conv_medium.e180_r384_in12k
+encoder_dim: 1280
+modality: trimodal
+audio_encoder_dim: 1280
+audio_finetune_last_n_stages: 0
+projection_hidden_dim: 1920
+projection_output_dim: 1280
+projection_dropout: 0.30
+signal_layout: raw
+batch_size: 4096
+max_epochs: 30
+learning_rate: 0.0012
+weight_decay: 0.0001
+warmup_fraction: 0.05
+grad_clip_norm: 1.0
+gradient_accumulation_steps: 1
+matryoshka_dims: [1280, 768, 512, 256, 128]
+matryoshka_weights: [1.0, 1.0, 1.0, 1.0, 1.0]
+loss_type: infonce
+temperature: 0.07
+temperature_min: 0.01
+learn_temperature: true
+hard_neg_k: 0
+hard_neg_weight: 2.0
+false_neg_threshold: 0.0
+feature_noise_std: 0.0
+feature_mask_ratio: 0.0
+mixup_alpha: 0.0
+num_workers: 8
+pin_memory: true
+prefetch_factor: 2
+persistent_workers: true
+mixed_precision: bf16
+checkpoint_dir: checkpoints
+save_every_n_epochs: 2
+early_stopping_patience: 8
+log_dir: runs
+benchmark_eval_every_epochs: 1

parameter_breakdown.json ADDED Viewed

	@@ -0,0 +1,9 @@

+{
+  "text_encoder": 22861056,
+  "image_encoder": 8434512,
+  "audio_encoder_native_mn20_merged": 19886566,
+  "image_projection": 12306560,
+  "audio_projection": 12306560,
+  "text_projection": 11323520,
+  "total_exact_loaded_params": 87118774
+}