Update README.md
Browse files
README.md
CHANGED
|
@@ -37,6 +37,22 @@ Each sample in MeowBench is structured as a **Multiple Choice Question (MCQ)**:
|
|
| 37 |
2. **Intent-Matched Synthesis:** Due to the scarcity of naturally synchronized quad-modal data, samples were synthesized by matching unimodal data sharing the same intent.
|
| 38 |
3. **Ethologist Audit:** To ensure biological and biomechanical plausibility, **eight Professional Feline Ethologists** manually reviewed every sample. Only samples where the biometric acceleration and heart-rate patterns were consistent with the audio-visual displays were retained.
|
| 39 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
# 🔗 The Meow-Omni Ecosystem
|
| 41 |
|
| 42 |
To facilitate reproducibility and further research in computational ethology, we have released the following components:
|
|
|
|
| 37 |
2. **Intent-Matched Synthesis:** Due to the scarcity of naturally synchronized quad-modal data, samples were synthesized by matching unimodal data sharing the same intent.
|
| 38 |
3. **Ethologist Audit:** To ensure biological and biomechanical plausibility, **eight Professional Feline Ethologists** manually reviewed every sample. Only samples where the biometric acceleration and heart-rate patterns were consistent with the audio-visual displays were retained.
|
| 39 |
|
| 40 |
+
# Leaderboard
|
| 41 |
+
|
| 42 |
+
| Model | Vision | Audio | TS | Accuracy |
|
| 43 |
+
|---|:---:|:---:|:---:|:---:|
|
| 44 |
+
| Acoustic SOTA (Ntalampiras et al., SVM/HMM) | | ✓ | | 36.86% |
|
| 45 |
+
| TS SOTA (Chen et al., 1D-CNN + LSTM on IMU) | | | ✓ | 48.98% |
|
| 46 |
+
| Video SOTA (Qwen3.5-122B-A10B, zero-shot) | ✓ | | | 61.95% |
|
| 47 |
+
| Qwen3.5-Omni-Plus (V + A) | ✓ | ✓ | | 65.36% |
|
| 48 |
+
| Qwen3.5-Omni-Plus (V + TS†) | ✓ | | ✓† | 66.21% |
|
| 49 |
+
| Qwen3.5-Omni-Plus (TS† + A) | | ✓ | ✓† | 42.15% |
|
| 50 |
+
| Qwen3.5-Omni-Plus (V + A + TS†) | ✓ | ✓ | ✓† | 66.89% |
|
| 51 |
+
| **Meow-Omni 1 (Ours)** | ✓ | ✓ | ✓ | **71.16%** |
|
| 52 |
+
|
| 53 |
+
† Qwen3.5-Omni-Plus does not accept raw time-series as a native modality; TS data was injected as a structured textual summary (array statistics per channel). Meow-Omni 1 processes raw TS natively.
|
| 54 |
+
|
| 55 |
+
|
| 56 |
# 🔗 The Meow-Omni Ecosystem
|
| 57 |
|
| 58 |
To facilitate reproducibility and further research in computational ethology, we have released the following components:
|