v0.1

.gitattributes

@@ -0,0 +1,38 @@
+*.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
+figures/paperfig-arch.png filter=lfs diff=lfs merge=lfs -text
+figures/results-table.pdf filter=lfs diff=lfs merge=lfs -text
+figures/results-table.png filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Zencastr
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

LICENSE-EAT

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 Wenxi Chen
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,169 @@
+---
+license: mit
+library_name: transformers
+pipeline_tag: audio-classification
+base_model: worstchan/EAT-base_epoch30_finetune_AS2M
+base_model_relation: finetune
+language:
+- multilingual
+tags:
+- audio
+- sound-event-detection
+- laughter-detection
+- podcast
+- onnx
+metrics:
+- average_precision
+---
+
+# Zencastr Laughter Detection
+
+Detect laughter events in close-microphone conversational speech. Built at [Zencastr](https://zencastr.com) for podcast audio and open-sourced as a tool for creators.
+## Usage
+
+Whole audio in, events out. Input can use any format ffmpeg can decode (path, URL, bytes, array) at any length or sample rate.
+```python
+from transformers import pipeline
+
+pipe = pipeline(
+    "laughter-detection",
+    model="zencastr/laughter-detection",
+    trust_remote_code=True,
+    device=0,
+)
+
+pipe("episode.mp3")
+# {
+#   "duration_s": 3591.4,
+#   "events": [
+#     {
+#       "start": 812.16,
+#       "end": 814.4,
+#       "score": 0.997,
+#     },
+#     ...
+#   ],
+#   "operating_point": {
+#     "name": "high_precision",
+#     "threshold": 0.99,
+#     "min_duration_s": 0.2,
+#     "merge_gap_s": 1.0
+#   },
+# }
+```
+
+From the model repository, run the command-line scripts with `uv`:
+
+```bash
+uv run events.py episode.mp3 --operating-point max_f1 --pretty  # PyTorch
+uv run events_onnx.py episode.mp3                               # ONNX
+```
+
+For clip-level classification of audio up to 4 s:
+
+```python
+clf = pipeline(
+    "audio-classification",
+    model="zencastr/laughter-detection",
+    trust_remote_code=True,
+)
+clf("clip.wav")
+# [
+#   {"label": "laughter", "score": 0.93},
+#   {"label": "no_laughter", "score": 0.07}
+# ]
+```
+
+## Inference
+
+<img src="figures/inference_simple.svg" alt="Inference architecture" width="720">
+
+Audio **x** is converted to a 128-bin log-mel spectrogram **X** at 16 kHz using 25 ms Hann windows and a 10 ms hop. The encoder **f** and laughter heads **g** then produce frame and clip probabilities. **φ** denotes the LoRA adapters. The model produces frame probabilities every 160 ms from 4 s windows evaluated with a 2 s hop. Scores from overlapping windows are combined by taking their maximum, then converted into events with approximate timestamps using a threshold, minimum duration, and merge gap. View the frame probabilities and tune the event parameters as follows:
+
+```python
+pipe("episode.mp3", threshold=0.9, return_frame_probs=True)
+```
+
+The `audio-classification` pipeline instead returns one attention-pooled probability for the entire clip.
+
+## Operating points
+
+Operating points are named presets for decoding frame probabilities into events.
+
+```python
+pipe("episode.mp3", operating_point="max_f1")
+```
+
+The presets were selected on ~17 h of internal podcast audio from 21 recordings using a 0.2 s matching tolerance. `high_precision` maximizes recall at precision ≥ 0.88 and is recommended when false alarms are costly; `high_recall` reaches recall 0.88 and is recommended when missed laughter is costly; `max_f1` maximizes F1 and provides a balance between precision and recall.
+
+| Preset | Threshold | Min duration | Merge gap | Precision | Recall | F1 | Onset error |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| `high_precision` (default) | 0.99 | 0.2 s | 1.0 s | 0.88 | 0.70 | 0.78 | 1.6 s |
+| `max_f1` | 0.97 | 0.16 s | 1.0 s | 0.80 | 0.77 | 0.78 | 1.6 s |
+| `high_recall` | 0.69 | 0.2 s | 0.16 s | 0.59 | 0.88 | 0.70 | 1.4 s |
+
+## Evaluation
+
+Our model achieves the highest mean clip-level average precision (AP) on three conversational benchmarks while maintaining a reasonable false positive rate (FPR) on other non-speech vocalizations.
+
+| Model | Podcast holdout AP ↑ | PodcastFillers AP ↑ | AMI AP ↑ | VocalSound AP ↑ | FSD50K FPR ↓ |
+|---|---:|---:|---:|---:|---:|
+| **Ours** | **0.973 ± 0.002** | 0.833 ± 0.011 | **0.948 ± 0.003** | 0.955 ± 0.005 | 15 ± 1% |
+| EAT | 0.948 ± 0.018 | 0.784 ± 0.049 | 0.923 ± 0.024 | 0.957 ± 0.013 | 8 ± 3% |
+| SenseVoice-S | 0.967 ± 0.012 | 0.771 ± 0.044 | 0.898 ± 0.030 | 0.935 ± 0.017 | 17 ± 4% |
+| CED-base | 0.950 ± 0.019 | 0.811 ± 0.060 | 0.854 ± 0.039 | 0.970 ± 0.010 | 4 ± 2% |
+| CED-tiny | 0.933 ± 0.024 | 0.798 ± 0.066 | 0.856 ± 0.032 | 0.956 ± 0.012 | 7 ± 2% |
+| BEATs | 0.908 ± 0.023 | 0.748 ± 0.064 | 0.842 ± 0.030 | 0.927 ± 0.017 | 15 ± 4% |
+| Gillick et al. | 0.951 ± 0.019 | 0.819 ± 0.042 | 0.853 ± 0.031 | 0.717 ± 0.031 | 48 ± 5% |
+
+Our row reports mean ± SD across multiple seeds while baselines rows report the performance of a single checkpoint ± the 95% bootstrap half-width. FSD50K FPR measures false positives on 1,136 clips of non-laughter vocalizations (coughing, sneezing, breathing, gasping, sighing, crying, and burping) at model-specific thresholds yielding 90% recall. CED, EAT, and BEATs use their AudioSet heads; SenseVoice uses its utterance-level laughter tag.
+
+## Inference speed
+
+Real-time factors measured on an NVIDIA L4 using the same 1,800 s of audio, batch 64, and the median of nine passes. All models use fp16 and PyTorch except ours, which uses ONNX opset 23 with fused attention. Timings include windowing and host-to-device transfer, and exclude model loading, audio decoding, and event decoding.
+
+| Model | Parameters | Window / hop | × realtime ↑ |
+|---|---:|---:|---:|
+| CED-tiny | 5.5 M | 4 s / 2 s | 4,540× |
+| CED-base | 85.7 M | 4 s / 2 s | 2,145× |
+| **Ours (ONNX)** | 90.7 M | 4 s / 2 s | **1,355×** |
+| EAT (PyTorch) | 90.4 M | 4 s / 2 s | 1,032× |
+| BEATs | 90.7 M | 2 s / 1 s | 410× |
+| SenseVoice-Small | 234.0 M | 4 s / 2 s | 328× |
+| Gillick et al. | 0.8 M | 1 s / 23 ms | 46× |
+
+
+## Files
+
+| File | Purpose |
+|---|---|
+| `model.safetensors` | PyTorch weights with LoRA merged and the AudioSet teacher head removed. Loads through `AutoModelForAudioClassification` / `AutoModelForAudioFrameClassification`. |
+| `model.onnx` | Float32, opset 17. Takes 16 kHz waveforms, including the mel front end in the graph. Runs with onnxruntime + numpy on CPU or GPU. |
+| `model_fp16.onnx` | Float16 encoder, opset 17, for GPU inference. |
+| `model_fp16_opset23.onnx` | Float16 with native ONNX Attention; onnxruntime ≥ 1.22, GPU. Used for the throughput result above. |
+| `export.json` | ONNX input/output contracts and parity checks against PyTorch. |
+| `laughter_core.py` | Shared numpy windowing, stitching, and event decoding. |
+
+## Provenance and licenses
+
+Model and code: MIT ([LICENSE](LICENSE)). Base encoder: [`worstchan/EAT-base_epoch30_finetune_AS2M`](https://huggingface.co/worstchan/EAT-base_epoch30_finetune_AS2M), revision `60d61e8b`, weights SHA-256 `5fa59385…69fff1b`; MIT, © 2024 Wenxi Chen ([LICENSE-EAT](LICENSE-EAT)). The base encoder was pretrained and fine-tuned on AudioSet.
+
+Adaptation used Zencastr platform audio collected under terms permitting model training. Public datasets were used only for evaluation and monitoring during adaptation. Dataset licenses: AudioSet annotations (CC BY 4.0), AMI (CC BY 4.0), VocalSound (CC BY-SA 4.0), PodcastFillers (CC BY-NC), and FSD50K (per-clip CC licenses).
+
+## Citation
+
+```bibtex
+@software{zencastr2026laughter,
+  title  = {Zencastr Laughter Detection},
+  author = {{Zencastr}},
+  year   = {2026},
+  url    = {https://huggingface.co/zencastr/laughter-detection}
+}
+
+@article{chen2024eat,
+  title   = {EAT: Self-Supervised Pre-Training with Efficient Audio Transformer},
+  author  = {Chen, Wenxi and Liang, Yuzhe and Ma, Ziyang and Zheng, Zhisheng and Chen, Xie},
+  journal = {arXiv preprint arXiv:2401.03497},
+  year    = {2024}
+}
+```

config.json

@@ -0,0 +1,191 @@
+{
+ "model_type": "eat-laughter",
+ "architectures": [
+  "LaughterForAudioFrameClassification"
+ ],
+ "auto_map": {
+  "AutoConfig": "configuration_laughter.LaughterConfig",
+  "AutoModel": "modeling_laughter.LaughterModel",
+  "AutoModelForAudioClassification": "modeling_laughter.LaughterForAudioClassification",
+  "AutoModelForAudioFrameClassification": "modeling_laughter.LaughterForAudioFrameClassification"
+ },
+ "custom_pipelines": {
+  "laughter-detection": {
+   "impl": "pipeline_laughter.LaughterDetectionPipeline",
+   "pt": [
+    "AutoModelForAudioFrameClassification"
+   ],
+   "tf": [],
+   "type": "audio"
+  }
+ },
+ "hidden_size": 768,
+ "num_hidden_layers": 12,
+ "num_attention_heads": 12,
+ "mlp_ratio": 4.0,
+ "patch_size": 16,
+ "num_mel_bins": 128,
+ "max_time_patches": 768,
+ "layer_norm_eps": 1e-06,
+ "qkv_bias": true,
+ "sample_rate": 16000,
+ "mel_frame_shift_s": 0.01,
+ "window_s": 4.0,
+ "window_hop_s": 2.0,
+ "default_operating_point": "high_precision",
+ "operating_points": {
+  "high_precision": {
+   "threshold": 0.99,
+   "min_duration_s": 0.2,
+   "merge_gap_s": 1.0
+  },
+  "high_recall": {
+   "threshold": 0.69,
+   "min_duration_s": 0.2,
+   "merge_gap_s": 0.16
+  },
+  "max_f1": {
+   "threshold": 0.97,
+   "min_duration_s": 0.16,
+   "merge_gap_s": 1.0
+  }
+ },
+ "id2label": {
+  "0": "no_laughter",
+  "1": "laughter"
+ },
+ "label2id": {
+  "no_laughter": 0,
+  "laughter": 1
+ },
+ "torch_dtype": "float32",
+ "release": {
+  "_comment": "v0.1 - SELECTED ARTIFACT is seed 4 of the 15-seed release family keep1-tnw4-lora-eat (the v0.1 LoRA recipe with the base checkpoint as teacher: keep-nsv gradient projection and teacher-weighted negatives). Selection rule, written before looking at test surfaces for this family: among seeds with validation AP >= 0.980 and event recall at the precision-first operating point in the family's upper band (>= 0.69; three seeds fell below on both criteria), the seed with the best event F1 at that point whose AP on every surface lies within 1.5 family standard deviations of the family mean. Every value below was read from the built model or the run's own artifacts.",
+  "version": "0.1.0",
+  "encoder": {
+   "base_model": "worstchan/EAT-base_epoch30_finetune_AS2M",
+   "base_model_revision": "60d61e8b2e9e5ba3be6860285de80cb7d625ccbb",
+   "base_model_weights_sha256": "5fa59385650eeb7b6c74414461306e055d943523cd7206c181e2d329969fff1b",
+   "base_model_license": "mit",
+   "_base_model_note": "The revision and hash pins are functional, not decoration: this repo is loaded with trust_remote_code, so an unpinned reference would execute whatever its author last pushed. Licence chain verified against the recorded provenance: HF metadata MIT + the upstream GitHub MIT LICENSE ((c) 2024 Wenxi Chen, read 2026-08-31, mirror confirmed by the authors' own README link), and the base checkpoint's AudioSet-2M fine-tuning data is CC BY 4.0 with attribution REQUIRED - carried in the card's Attribution section.",
+   "frozen": true,
+   "lora": {
+    "rank": 8,
+    "alpha": 16,
+    "dropout": 0.05,
+    "target_modules": [
+     "qkv"
+    ],
+    "injected_layers": 12,
+    "scaling_alpha_over_rank": 2.0,
+    "trainable_params": {
+     "head": 1538,
+     "encoder_lora": 294912,
+     "total": 296450
+    }
+   },
+   "embed_dim_verify": 768,
+   "depth_verify": 12,
+   "num_heads_verify": 12
+  },
+  "audio": {
+   "sample_rate": 16000,
+   "mono": true,
+   "n_mels_verify": 128,
+   "win_length_verify": 400,
+   "hop_length_verify": 160,
+   "normalize": "(x - -4.268) / (4.569 * 2)",
+   "_features": "torchaudio.compliance.kaldi.fbank, htk_compat=True, use_energy=False, window_type='hanning', num_mel_bins=128, dither=0.0, frame_shift=10 — then the normalize above. Front-end is Python, deliberately outside the ONNX graph."
+  },
+  "head": {
+   "type": "frame_logits_attention_pooling",
+   "frame_hop_s": 0.16,
+   "pooling": "attention",
+   "num_classes": 1,
+   "_frame_hop_s_derivation": "16 mel frames per patch x 10 ms hop; measured, not assumed"
+  },
+  "inference": {
+   "window_s": 4.0,
+   "window_hop_s": 2.0,
+   "min_input_s": 0.175,
+   "pad_short_inputs": true,
+   "frame_prob_stitching": "max_over_overlapping_windows",
+   "_decoder_note": "min_duration and merge_gap are per operating point, not global - calibration selects a different pair for each (high_precision favours longer minimum duration and wider merging, high_recall the opposite). See operating_points below.",
+   "event_score": "peak_frame_probability",
+   "_min_input_s_derivation": "2800 samples at 16 kHz — the shortest input that runs, found by sweeping in 10 ms steps. Matches the analytic bound: 400 window samples + 15 x 160 hop = 16 mel frames = one patch."
+  },
+  "calibration": {
+   "clip_level": null,
+   "_note": "No clip-level calibration ships. Clip probabilities are the MIL-pooled head output, uncalibrated. Fitting a temperature or isotonic map on val would inherit val's duration leak, so the shipped operating points are event-level only — see operating_points._clip_level_not_shipped."
+  },
+  "operating_points": {
+   "default": "event_high_precision",
+   "_level_note": "clip_* threshold MIL-pooled clip probabilities; event_* threshold frame probabilities and decode to events. Different quantities on different scales — never substitute one for the other.",
+   "event_high_precision": {
+    "threshold": 0.99,
+    "min_duration_s": 0.2,
+    "merge_gap_s": 1.0,
+    "precision": 0.881,
+    "recall": 0.7,
+    "event_f1": 0.78,
+    "onset_mae_s": 1.61
+   },
+   "event_high_recall": {
+    "threshold": 0.69,
+    "min_duration_s": 0.2,
+    "merge_gap_s": 0.16,
+    "precision": 0.587,
+    "recall": 0.88,
+    "event_f1": 0.704,
+    "onset_mae_s": 1.35,
+    "_grid_limited": "min_duration_s sits at the smallest value swept, so a better setting may exist below 0.16 s — this is the best of what was looked at."
+   },
+   "_clip_level_not_shipped": "Deliberately absent. Clip-level thresholds are fitted on validation data, where clip duration leaks the label (duration alone scores AP 0.566 vs a 0.378 base rate), and that cue does not exist at inference. Use the event operating points.",
+   "_calibrated_on": "16.85 h of internal audit audio, 3276 decoder settings swept jointly (threshold x min_duration x merge_gap). Tolerance 0.2 s. high_precision is the highest-recall setting with precision >= 0.88, high_recall the setting reaching recall 0.88, max_f1 the best F1, all against the corpus's labelled laughs (precision is a lower bound). Meetings figures come from AMI.",
+   "event_max_f1": {
+    "threshold": 0.97,
+    "min_duration_s": 0.16,
+    "merge_gap_s": 1.0,
+    "precision": 0.798,
+    "recall": 0.766,
+    "event_f1": 0.782,
+    "onset_mae_s": 1.55
+   }
+  },
+  "release_contents": {
+   "_note": "These files back the card's claim that the significance analyses can be rerun without a GPU.",
+   "eval/probs/": "per-clip probabilities, one .npy per (surface, model) - 48 files, public surfaces only",
+   "eval/<surface>.clips.tsv": "clip metadata (id, target, class, recording, language, duration) for the 3 public surfaces",
+   "eval/MANIFEST.json": "bundle provenance: family, members, bootstrap params, code revision",
+   "eval/configs/": "the family's shared training config, shipped once as laughter-v0.1.json (the 15 runs are seeds 0-14 of it)",
+   "_internal_surface": "annotation-eval-v5 (2,297 clips, 21 recordings; digest 2dcdba95c052bbde), the rebuilt internal holdout in which each laughter clip is cut from the annotated speaker's track. Per-clip data for it is not shipped."
+  },
+  "provenance": {
+   "train_data_hash": "d63ba2795f597baa",
+   "training_commit": "e4db7af66b151e230c6075e28253302b9b48ff6e",
+   "_train_data_hash_note": "sha256-derived digest over the training-determining columns of the training set (9600 rows, 6694/1366/1540 train/val/test). Recomputed on the set the run trained from; equal to the digest of the original labeled-v8v9 build.",
+   "_training_commit_note": "Recorded by train.py at run time. The tree was dirty at that commit: the uncommitted files were training/hf_convert.py (accept a full-fine-tune checkpoint) and the hq5a job scripts, none of them on the training path."
+  },
+  "total_params": 89974274,
+  "_total_params_note": "Tensors in model.safetensors: the encoder with LoRA folded in plus the 1,538-parameter laughter head. The training checkpoint held 90,675,985, the extra being the unmerged LoRA factors and the AudioSet classifier, neither of which ships.",
+  "_open": [],
+  "training": {
+   "recipe": {
+    "adapt": "lora",
+    "epochs": 60,
+    "patience": 8,
+    "batch_size": 32,
+    "head_lr": 0.001,
+    "encoder_lr": 1e-05,
+    "keep_nsv": 1.0,
+    "teacher_neg_weight": 4.0,
+    "seed": 4,
+    "family": "keep1-tnw4-lora-eat",
+    "family_size": 15
+   },
+   "teacher": "The base checkpoint itself, read through its own AudioSet classifier (the student with adapters off; no second model, no additional data). Two uses: (1) a non-laughter clip's loss is weighted 1 + 4 x the teacher's strongest posterior over cough, sneeze, throat clearing, sniff, breathing, gasp, sigh, burp and hiccup; (2) the component of each encoder step that would raise a BCE between the student's and the teacher's AudioSet logits on 13 non-speech classes is removed (A-GEM projection, alpha 1). Measured effect vs the same recipe without the teacher: laughter AP unchanged on every surface; laugh-vs-cough AUROC 0.939 -> 0.956 on VocalSound and 0.861 -> 0.876 on AMI; FSD50K non-laughter vocalisations called laughter at 90% laughter recall 19% -> 15%.",
+   "monitoring_only": "An FSD50K panel (4,269 dev clips with cough/sneeze/breathing/crying/... labels, CC licences mixed) is scored every epoch through the frozen AudioSet head to detect forgetting. No FSD50K audio is trained on; the trainer refuses to start if a panel clip is also a training clip.",
+   "training_data": "9,600 internal clips (labeled-v8v9), digest below. No public dataset was used for training."
+  }
+ }
+}

configuration_laughter.py

@@ -0,0 +1,50 @@
+"""Configuration for the zencastr/laughter-detection model (EAT-base encoder + MIL laughter head)."""
+from transformers import PreTrainedConfig
+
+
+class LaughterConfig(PreTrainedConfig):
+    model_type = "eat-laughter"
+
+    def __init__(
+        self,
+        hidden_size: int = 768,
+        num_hidden_layers: int = 12,
+        num_attention_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        patch_size: int = 16,
+        num_mel_bins: int = 128,
+        max_time_patches: int = 768,
+        layer_norm_eps: float = 1e-6,
+        qkv_bias: bool = True,
+        sample_rate: int = 16000,
+        mel_frame_shift_s: float = 0.01,
+        window_s: float = 4.0,
+        window_hop_s: float = 2.0,
+        default_operating_point: str = "high_precision",
+        operating_points: dict | None = None,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.mlp_ratio = mlp_ratio
+        self.patch_size = patch_size
+        self.num_mel_bins = num_mel_bins
+        self.max_time_patches = max_time_patches
+        self.layer_norm_eps = layer_norm_eps
+        self.qkv_bias = qkv_bias
+        self.sample_rate = sample_rate
+        self.mel_frame_shift_s = mel_frame_shift_s
+        self.window_s = window_s
+        self.window_hop_s = window_hop_s
+        self.default_operating_point = default_operating_point
+        # threshold / min_duration_s / merge_gap_s per named point; the decoder needs all three
+        self.operating_points = operating_points or {}
+        kwargs.setdefault("id2label", {0: "no_laughter", 1: "laughter"})
+        kwargs.setdefault("label2id", {"no_laughter": 0, "laughter": 1})
+        super().__init__(**kwargs)
+
+    @property
+    def frame_hop_s(self) -> float:
+        """One output frame per time patch: patch_size mel frames x the mel frame shift (0.16 s)."""
+        return self.patch_size * self.mel_frame_shift_s

eval/MANIFEST.json

@@ -0,0 +1,554 @@
+{
+ "bootstrap": {
+  "n": 2000,
+  "seed": 0,
+  "unit": "recording"
+ },
+ "code_dirty": [
+  "hf-release-v0.1/config.json",
+  "training/hf_convert.py",
+  "training/hq5a_cough_eval.sbatch",
+  "training/hq5a_evalsets.sbatch",
+  "training/release_gate.py",
+  "hf-release-v0.1/release-manifest.json",
+  "training/hq5a_event_calib.sbatch"
+ ],
+ "code_revision": "e4db7af66b151e230c6075e28253302b9b48ff6e",
+ "family": "laughter-v0.1",
+ "members": [
+  "laughter-v0.1-seed0",
+  "laughter-v0.1-seed1",
+  "laughter-v0.1-seed10",
+  "laughter-v0.1-seed11",
+  "laughter-v0.1-seed12",
+  "laughter-v0.1-seed13",
+  "laughter-v0.1-seed14",
+  "laughter-v0.1-seed2",
+  "laughter-v0.1-seed3",
+  "laughter-v0.1-seed4",
+  "laughter-v0.1-seed5",
+  "laughter-v0.1-seed6",
+  "laughter-v0.1-seed7",
+  "laughter-v0.1-seed8",
+  "laughter-v0.1-seed9"
+ ],
+ "reproduce": "For each public surface: AP of each probs/<surface>__laughter-v0.1-seedN.npy against target>=0.5 in <surface>.clips.tsv, mean over the 15 seeds = family_mean; margin = family_mean - reference AP; CI by paired recording (speaker on VocalSound) bootstrap, 2000 resamples, seed 0.",
+ "rule": "internal-and-any-public",
+ "set_digests": {
+  "ami-eval-v1": "1ab59b11012d18d9",
+  "annotation-eval-v5": "2dcdba95c052bbde",
+  "podcastfillers-eval": "e4deca38c6380c35",
+  "public-vocalsound": "c4557271c7f68f70",
+  "fsd50k-panel": "b55cd8789c24fda8"
+ },
+ "table": {
+  "ami-eval-v1": {
+   "ci_hi": 0.03942520109963582,
+   "ci_lo": 0.012154065156311903,
+   "family_mean": 0.9483739882117678,
+   "floor_ci_upper": 0.0039598598827272615,
+   "floor_sigma": 0.002556051282820607,
+   "label": "AMI",
+   "margin": 0.02538138623377517,
+   "member_ap": [
+    0.9507902186015077,
+    0.9500519992061698,
+    0.9452210675451327,
+    0.9503056961627866,
+    0.9508894430969833,
+    0.9470331036343588,
+    0.9450138166749027,
+    0.9465480039960334,
+    0.9513989118555961,
+    0.9483455133944823,
+    0.9432426745514125,
+    0.9511137070448046,
+    0.9483085823703894,
+    0.9485920621237138,
+    0.9487550229182443
+   ],
+   "n": 15,
+   "n_clips": 3082,
+   "n_recordings": 24,
+   "reference_ap": 0.9229926019779926,
+   "reference_model": "eat"
+  },
+  "annotation-eval-v5": {
+   "ci_hi": 0.023578607063105828,
+   "ci_lo": -0.014470678820449834,
+   "family_mean": 0.8996943593326668,
+   "floor_ci_upper": 0.004720526387522345,
+   "floor_sigma": 0.003047054159932776,
+   "label": "internal",
+   "margin": 0.0028852774625475197,
+   "member_ap": [
+    0.8975475415661825,
+    0.9039947028281456,
+    0.8983932783475506,
+    0.8993346068750109,
+    0.9005183386530095,
+    0.9016603462847035,
+    0.9028680945100267,
+    0.8955774413407963,
+    0.9002877424905875,
+    0.9031921793787044,
+    0.8952398854096171,
+    0.9011704571931498,
+    0.8995806005577214,
+    0.8942340859359875,
+    0.9018160886188078
+   ],
+   "n": 15,
+   "n_clips": 2297,
+   "n_recordings": 21,
+   "reference_ap": 0.8968090818701193,
+   "reference_model": "ced-base",
+   "_per_clip": "not shipped (internal audio)"
+  },
+  "podcastfillers-eval": {
+   "ci_hi": 0.050521526766975876,
+   "ci_lo": -0.02075922278013193,
+   "family_mean": 0.8333527146573345,
+   "floor_ci_upper": 0.01673321979650875,
+   "floor_sigma": 0.010801131654468506,
+   "label": "PF",
+   "margin": 0.014778337531855201,
+   "member_ap": [
+    0.8344361951641851,
+    0.841598928384278,
+    0.8411140826012734,
+    0.8366294402860901,
+    0.8337691285784827,
+    0.8382277485746962,
+    0.8403275294236768,
+    0.831845088772049,
+    0.8383355800012807,
+    0.8350686755602265,
+    0.7975593280141196,
+    0.8376021852481,
+    0.8343551388336643,
+    0.8279070269767151,
+    0.8315146434411776
+   ],
+   "n": 15,
+   "n_clips": 10398,
+   "n_recordings": 20,
+   "reference_ap": 0.8185743771254793,
+   "reference_model": "gillick"
+  },
+  "public-vocalsound": {
+   "ci_hi": -0.005603668917087099,
+   "ci_lo": -0.02523479335876062,
+   "family_mean": 0.9550608658482184,
+   "floor_ci_upper": 0.007326952674220729,
+   "floor_sigma": 0.004729477137259021,
+   "label": "VS",
+   "margin": -0.015235673719981468,
+   "member_ap": [
+    0.9520376579742674,
+    0.9552183790146299,
+    0.9540970381377161,
+    0.9509495673970976,
+    0.9519229996151687,
+    0.9537347287207015,
+    0.9557549955025102,
+    0.9619265200931599,
+    0.9544285936941963,
+    0.9517720709172127,
+    0.9652754551778516,
+    0.9524594157733894,
+    0.9488156226153927,
+    0.9624520243219058,
+    0.9550679187680756
+   ],
+   "n": 15,
+   "n_clips": 3591,
+   "n_recordings": 574,
+   "reference_ap": 0.9702965395681998,
+   "reference_model": "ced-base"
+  }
+ },
+ "tolerance": {
+  "ap": 1e-12,
+  "ci": 0.005
+ },
+ "verdict": "FAIL: internal",
+ "z_whole_ci": {
+  "AMI": 6.409667762358887,
+  "PF": 0.8831735739787857,
+  "VS": -2.0794011367900485,
+  "internal": 0.6112194331068893
+ },
+ "curation": {
+  "by": "release packaging",
+  "date": "2026-09-05",
+  "note": "Public bundle, curated from the internal eval bundle: the private internal surface's per-clip data is excluded (its digest and aggregate row are kept so the card's internal row stays auditable), run names are mapped to release naming (laughter-v0.1-seed0..14, order preserved against the member_ap arrays), the fifteen per-seed configs ship once as configs/laughter-v0.1.json (runs are seeds 0-14 of it), and operational fields are removed from that config. eval/podcastfillers-relabel.tsv records the 100 PodcastFillers non-laughter clips both model families scored above 0.9 and the listening verdict on each (97 contain audible laughter); the table above uses the published labels. 2026-09-09: the FSD50K NSV panel added (fsd50k-panel.clips.tsv, probs for the 15 seeds and ced-base); its column is defined under fsd50k_panel. 2026-09-10: the two no-teacher arms retrained on seeds 0-14 so every trained row shares the same seeds; card_table holds every cell."
+ },
+ "fsd50k_panel": {
+  "label": "FSD50K NSV",
+  "what": "Share of FSD50K non-laughter vocalisation clips (any of cough, sneeze, breathing, gasp, sigh, crying, burping; no Laughter label) scored as laughter, at the threshold where the model recovers 90% of the panel's Laughter clips. Each clip's score is the max over its 4 s windows (hop 4 s). Lower is better.",
+  "n_laughter": 933,
+  "n_nsv": 1136,
+  "n_clips": 4269,
+  "recall": 0.9,
+  "family_mean": 0.1459507042253521,
+  "family_sd": 0.013560746885563212,
+  "member_nsv_fp": [
+   0.14964788732394366,
+   0.15316901408450703,
+   0.1267605633802817,
+   0.13556338028169015,
+   0.1637323943661972,
+   0.13468309859154928,
+   0.13556338028169015,
+   0.1619718309859155,
+   0.12147887323943662,
+   0.13908450704225353,
+   0.16549295774647887,
+   0.14964788732394366,
+   0.15933098591549297,
+   0.14788732394366197,
+   0.14524647887323944
+  ],
+  "reference_model": "ced-base",
+  "reference_nsv_fp": 0.03609154929577465,
+  "others": {
+   "eat-lora-no-teacher": {
+    "nsv_fp": 0.1931338028169014,
+    "sd": 0.01694497340594801,
+    "n": 15
+   },
+   "eat-full-finetune-teacher": {
+    "nsv_fp": 0.17951877934272303,
+    "sd": 0.04972447014311274,
+    "n": 15
+   },
+   "eat-full-finetune-no-teacher": {
+    "nsv_fp": 0.26085680751173707,
+    "sd": 0.030258187344898973,
+    "n": 15
+   },
+   "eat": {
+    "nsv_fp": 0.0801056338028169,
+    "n": 1
+   },
+   "ced-tiny": {
+    "nsv_fp": 0.07130281690140845,
+    "n": 1
+   },
+   "beats": {
+    "nsv_fp": 0.14524647887323944,
+    "n": 1
+   },
+   "gillick": {
+    "nsv_fp": 0.48327464788732394,
+    "n": 1
+   }
+  },
+  "reproduce": "s = probs/fsd50k-panel__<model>.npy aligned with fsd50k-panel.clips.tsv; thr = the 840th highest score among laughter==1 rows (ceil(0.9*933)); nsv_fp = mean(s[nsv==1] >= thr). Clips are FSD50K dev audio (train and val folds) that carry panel labels; none of it is training data."
+ },
+ "card_table": {
+  "note": "Every cell of the card's results table. Trained rows: mean and sd over seeds 0-14. Off-the-shelf rows: value and 95% bootstrap half-width (2000 draws, seed 0) over recordings (speakers on VocalSound, clips on the FSD50K panel). Built by training.card_results.",
+  "columns": [
+   "annotation-eval-v5",
+   "podcastfillers-eval",
+   "ami-eval-v1",
+   "public-vocalsound",
+   "fsd50k-panel"
+  ],
+  "panel_recall": 0.9,
+  "bootstrap": 2000,
+  "n_laughter": 933,
+  "n_nsv": 1136,
+  "rows": {
+   "**this release**: EAT + LoRA, teacher-trained": {
+    "annotation-eval-v5": {
+     "value": 0.8996943593326668,
+     "spread": 0.0029931705939927142,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "podcastfillers-eval": {
+     "value": 0.8333527146573345,
+     "spread": 0.010610126355847991,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "ami-eval-v1": {
+     "value": 0.9483739882117678,
+     "spread": 0.002510850524772026,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "public-vocalsound": {
+     "value": 0.9550608658482184,
+     "spread": 0.0046458419014504355,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "fsd50k-panel": {
+     "value": 0.14595070422535214,
+     "spread": 0.013560746885563212,
+     "spread_kind": "seed sd",
+     "n": 15
+    }
+   },
+   "EAT + LoRA, no teacher": {
+    "annotation-eval-v5": {
+     "value": 0.9022181395955502,
+     "spread": 0.0028596226591157932,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "podcastfillers-eval": {
+     "value": 0.846728733967997,
+     "spread": 0.014180319446455675,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "ami-eval-v1": {
+     "value": 0.9524466643323456,
+     "spread": 0.003276968221102732,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "public-vocalsound": {
+     "value": 0.9420809875492873,
+     "spread": 0.006998231467778743,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "fsd50k-panel": {
+     "value": 0.1931338028169014,
+     "spread": 0.01694497340594801,
+     "spread_kind": "seed sd",
+     "n": 15
+    }
+   },
+   "EAT full fine-tune, teacher-trained": {
+    "annotation-eval-v5": {
+     "value": 0.9166429554840164,
+     "spread": 0.005625359041947865,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "podcastfillers-eval": {
+     "value": 0.8750703685982714,
+     "spread": 0.011921685111299707,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "ami-eval-v1": {
+     "value": 0.9446209554473263,
+     "spread": 0.005212745805492586,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "public-vocalsound": {
+     "value": 0.9437988023703398,
+     "spread": 0.01700459450721973,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "fsd50k-panel": {
+     "value": 0.17951877934272303,
+     "spread": 0.04972447014311274,
+     "spread_kind": "seed sd",
+     "n": 15
+    }
+   },
+   "EAT full fine-tune, no teacher": {
+    "annotation-eval-v5": {
+     "value": 0.918610648846479,
+     "spread": 0.006374031205233863,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "podcastfillers-eval": {
+     "value": 0.8808029276378652,
+     "spread": 0.008482409285809667,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "ami-eval-v1": {
+     "value": 0.9523479177219377,
+     "spread": 0.004569827830633127,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "public-vocalsound": {
+     "value": 0.9089073852987579,
+     "spread": 0.059372105425155815,
+     "spread_kind": "seed sd",
+     "n": 15
+    },
+    "fsd50k-panel": {
+     "value": 0.26085680751173707,
+     "spread": 0.030258187344898973,
+     "spread_kind": "seed sd",
+     "n": 15
+    }
+   },
+   "CED-base, AudioSet head": {
+    "annotation-eval-v5": {
+     "value": 0.8968090818701193,
+     "spread": 0.0228192185597908,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "podcastfillers-eval": {
+     "value": 0.8105831097976058,
+     "spread": 0.06036189372814932,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "ami-eval-v1": {
+     "value": 0.8541544974880243,
+     "spread": 0.039408864655789466,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "public-vocalsound": {
+     "value": 0.9702965395681998,
+     "spread": 0.009631731710918856,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "fsd50k-panel": {
+     "value": 0.03609154929577465,
+     "spread": 0.016285211267605633,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    }
+   },
+   "EAT, AudioSet head (the base checkpoint, zero-shot)": {
+    "annotation-eval-v5": {
+     "value": 0.8829045193537908,
+     "spread": 0.02558446961170535,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "podcastfillers-eval": {
+     "value": 0.7844658804345496,
+     "spread": 0.04947444610463847,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "ami-eval-v1": {
+     "value": 0.9229926019779926,
+     "spread": 0.024309727861856867,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "public-vocalsound": {
+     "value": 0.9571953064111093,
+     "spread": 0.013309956523687383,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "fsd50k-panel": {
+     "value": 0.0801056338028169,
+     "spread": 0.027739876760563317,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    }
+   },
+   "CED-tiny, AudioSet head": {
+    "annotation-eval-v5": {
+     "value": 0.8857709447136373,
+     "spread": 0.02259637836585593,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "podcastfillers-eval": {
+     "value": 0.7980219459500315,
+     "spread": 0.06555965664476748,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "ami-eval-v1": {
+     "value": 0.8564672033882755,
+     "spread": 0.03207273515356468,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "public-vocalsound": {
+     "value": 0.9558670210236383,
+     "spread": 0.012283620217255986,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "fsd50k-panel": {
+     "value": 0.07130281690140845,
+     "spread": 0.02112676056338028,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    }
+   },
+   "BEATs, AudioSet head": {
+    "annotation-eval-v5": {
+     "value": 0.8383485927709583,
+     "spread": 0.031978569016706926,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "podcastfillers-eval": {
+     "value": 0.7475556060918509,
+     "spread": 0.06449430520655708,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "ami-eval-v1": {
+     "value": 0.8419949943494651,
+     "spread": 0.029730828291308575,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "public-vocalsound": {
+     "value": 0.9273063305843975,
+     "spread": 0.016602341784197117,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "fsd50k-panel": {
+     "value": 0.14524647887323944,
+     "spread": 0.03962367957746479,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    }
+   },
+   "Gillick et al.": {
+    "annotation-eval-v5": {
+     "value": 0.8590035695842937,
+     "spread": 0.022847874353034203,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "podcastfillers-eval": {
+     "value": 0.8185743771254793,
+     "spread": 0.04195457050713697,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "ami-eval-v1": {
+     "value": 0.8525958378855036,
+     "spread": 0.030944159938566085,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "public-vocalsound": {
+     "value": 0.7171790828625292,
+     "spread": 0.0312806324260928,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    },
+    "fsd50k-panel": {
+     "value": 0.48327464788732394,
+     "spread": 0.053268045774647804,
+     "spread_kind": "bootstrap 95% half-width",
+     "n": 1
+    }
+   }
+  }
+ }
+}

eval/configs/laughter-v0.1.json

@@ -0,0 +1,27 @@
+{
+ "sets": [
+  "<internal labeled set - private>"
+ ],
+ "encoder": "eat",
+ "adapt": "lora",
+ "epochs": 60,
+ "batch_size": 32,
+ "head_lr": 0.001,
+ "encoder_lr": 1e-05,
+ "patience": 8,
+ "truncate_blocks": null,
+ "lr_t_max": null,
+ "train_fraction": 1.0,
+ "distill_cough": 0.0,
+ "no_val": false,
+ "save_every_epoch": false,
+ "nsv_head": 0.0,
+ "nsv_weight": 1.0,
+ "keep_nsv": 1.0,
+ "teacher_neg_weight": 4.0,
+ "teacher_aux": 0.0,
+ "nsv_panel": [
+  "<FSD50K NSV panel - monitoring only>"
+ ],
+ "init_from": null
+}

eval/podcastfillers-relabel.tsv

@@ -0,0 +1,101 @@
+clip_id	pf_label	episode	v0.1_score	candidate_score	verdict
+pf_30199	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.996	0.996	laughter
+pf_74068	breath	The Genealogy Happy Hour_Episode 20_Privacy	0.994	0.996	laughter
+pf_11944	speech	The RegenNarration_79. Gondwana Link The most exciting restoration at scale in Australia, with Keith Bradby OAM	0.995	0.996	laughter
+pf_80857	speech	Beer Tour Guy Podcast_Episode 12 - Mark Larson - Falling Down Beer Company	0.993	0.995	laughter
+pf_85103	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.991	0.993	laughter
+pf_20739	speech	Spaces Podcast_Evolution of Architecture I	0.992	0.993	laughter
+pf_50969	breath	Between You & Me Podcast_Ep 54_MORIAH SMALLBONE Becoming a Lead Lady	0.99	0.993	laughter
+pf_08802	speech	Spaces Podcast_Evolution of Architecture I	0.993	0.993	laughter
+pf_68449	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.992	0.993	laughter
+pf_05868	breath	Beer Tour Guy Podcast_Episode 12 - Mark Larson - Falling Down Beer Company	0.988	0.992	laughter
+pf_70216	speech	Spaces Podcast_Evolution of Architecture I	0.992	0.988	laughter
+pf_74417	speech	The Booktopia Books Podcast_John Birmingham on How to Be a Writer	0.99	0.992	laughter
+pf_29389	breath	Footy for Two_Transfers and Fantasy League	0.992	0.991	laughter
+pf_02945	speech	g1 pdcst_Foodie Call #11 The 2016 Digest	0.991	0.992	laughter
+pf_13808	speech	Break The Cycle_Episode 7 Patriarchy	0.991	0.991	laughter
+pf_31378	breath	The Genealogy Happy Hour_Episode 20_Privacy	0.99	0.991	laughter
+pf_13939	speech	Spaces Podcast_Evolution of Architecture I	0.988	0.99	laughter
+pf_20786	speech	Break The Cycle_Episode 7 Patriarchy	0.989	0.99	laughter
+pf_41961	breath	Break The Cycle_Episode 7 Patriarchy	0.988	0.99	laughter
+pf_85602	breath	The RegenNarration_79. Gondwana Link The most exciting restoration at scale in Australia, with Keith Bradby OAM	0.982	0.99	laughter
+pf_45598	speech	The RegenNarration_79. Gondwana Link The most exciting restoration at scale in Australia, with Keith Bradby OAM	0.979	0.989	laughter
+pf_56587	speech	Spaces Podcast_Evolution of Architecture I	0.988	0.987	laughter
+pf_15529	speech	The Booktopia Books Podcast_John Birmingham on How to Be a Writer	0.988	0.978	laughter
+pf_14343	other	Food as medicine_Integrative Medicine	0.988	0.984	laughter
+pf_77467	speech	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.979	0.987	laughter
+pf_19039	overlap	The Booktopia Books Podcast_John Birmingham on How to Be a Writer	0.985	0.987	laughter
+pf_19970	speech	Between You & Me Podcast_Ep 54_MORIAH SMALLBONE Becoming a Lead Lady	0.987	0.985	laughter
+pf_78788	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.987	0.981	laughter
+pf_01704	breath	Between You & Me Podcast_Ep 54_MORIAH SMALLBONE Becoming a Lead Lady	0.986	0.979	laughter
+pf_68606	breath	The RegenNarration_79. Gondwana Link The most exciting restoration at scale in Australia, with Keith Bradby OAM	0.979	0.986	laughter
+pf_09148	breath	Between You & Me Podcast_Ep 54_MORIAH SMALLBONE Becoming a Lead Lady	0.985	0.985	laughter
+pf_49544	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.985	0.971	laughter
+pf_82668	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.976	0.985	laughter
+pf_51275	breath	Break The Cycle_Episode 7 Patriarchy	0.984	0.98	laughter
+pf_48894	breath	Break The Cycle_Episode 7 Patriarchy	0.983	0.981	laughter
+pf_08149	speech	Spaces Podcast_Evolution of Architecture I	0.983	0.966	laughter
+pf_33722	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.975	0.982	laughter
+pf_27533	overlap	Food as medicine_Integrative Medicine	0.972	0.982	laughter
+pf_45870	breath	Tin Pan Alley_On the Road Charlotte Eriksson	0.982	0.979	laughter
+pf_28283	speech	The RegenNarration_79. Gondwana Link The most exciting restoration at scale in Australia, with Keith Bradby OAM	0.965	0.981	laughter
+pf_60849	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.978	0.981	laughter
+pf_23028	overlap	Free Podcast Day_Home Grown Girls – Apr 7 2018	0.981	0.978	laughter
+pf_41112	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.977	0.98	laughter
+pf_18485	breath	g1 pdcst_Foodie Call #11 The 2016 Digest	0.979	0.98	laughter
+pf_44027	event	Footy for Two_Transfers and Fantasy League	0.98	0.958	laughter
+pf_60970	speech	Spaces Podcast_Evolution of Architecture I	0.98	0.98	laughter
+pf_67669	speech	Beer Tour Guy Podcast_Episode 12 - Mark Larson - Falling Down Beer Company	0.98	0.976	laughter
+pf_36287	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.968	0.979	laughter
+pf_09295	breath	Food as medicine_Integrative Medicine	0.973	0.977	laughter
+pf_64782	speech	Spaces Podcast_Evolution of Architecture I	0.976	0.977	laughter
+pf_74733	speech	Spaces Podcast_Evolution of Architecture I	0.977	0.966	laughter
+pf_23984	speech	The Housing Hour_History of Coal Creek Miners	0.974	0.977	laughter
+pf_09557	breath	The Booktopia Books Podcast_John Birmingham on How to Be a Writer	0.962	0.974	laughter
+pf_23130	speech	Design Story_025 Moira Cullen Design is Soft Power	0.966	0.974	laughter
+pf_78502	speech	Beer Tour Guy Podcast_Episode 12 - Mark Larson - Falling Down Beer Company	0.973	0.973	laughter
+pf_43620	speech	The Housing Hour_History of Coal Creek Miners	0.972	0.973	laughter
+pf_69203	speech	g1 pdcst_Foodie Call #11 The 2016 Digest	0.963	0.972	laughter
+pf_66810	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.962	0.971	laughter
+pf_36989	breath	Break The Cycle_Episode 7 Patriarchy	0.97	0.96	laughter
+pf_18407	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.961	0.97	laughter
+pf_48234	speech	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.968	0.969	laughter
+pf_72082	speech	Beer Tour Guy Podcast_Episode 12 - Mark Larson - Falling Down Beer Company	0.966	0.969	laughter
+pf_55357	speech	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.943	0.968	laughter
+pf_35665	speech	Spaces Podcast_Evolution of Architecture I	0.968	0.958	laughter
+pf_70852	speech	Public Access America_Obama Foundation Town Hall Europe	0.962	0.965	laughter
+pf_07535	speech	g1 pdcst_Foodie Call #11 The 2016 Digest	0.95	0.963	laughter
+pf_18576	speech	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.946	0.959	laughter
+pf_51021	breath	g1 pdcst_Foodie Call #11 The 2016 Digest	0.939	0.957	laughter
+pf_19038	overlap	g1 pdcst_Foodie Call #11 The 2016 Digest	0.952	0.956	laughter
+pf_54944	breath	Between You & Me Podcast_Ep 54_MORIAH SMALLBONE Becoming a Lead Lady	0.954	0.942	laughter
+pf_62127	speech	The Housing Hour_History of Coal Creek Miners	0.921	0.952	laughter
+pf_50835	breath	Break The Cycle_Episode 7 Patriarchy	0.946	0.951	laughter
+pf_17226	speech	The RegenNarration_79. Gondwana Link The most exciting restoration at scale in Australia, with Keith Bradby OAM	0.951	0.941	laughter
+pf_41917	speech	Spaces Podcast_Evolution of Architecture I	0.951	0.949	laughter
+pf_41975	event	The Booktopia Books Podcast_John Birmingham on How to Be a Writer	0.946	0.951	laughter
+pf_70849	speech	Free Podcast Day_Home Grown Girls – Apr 7 2018	0.929	0.948	laughter
+pf_57530	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.943	0.947	laughter
+pf_55480	speech	The Housing Hour_History of Coal Creek Miners	0.909	0.947	laughter
+pf_42257	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.946	0.941	laughter
+pf_17471	breath	The Genealogy Happy Hour_Episode 20_Privacy	0.944	0.942	laughter
+pf_48051	breath	Break The Cycle_Episode 7 Patriarchy	0.943	0.943	laughter
+pf_41628	breath	Between You & Me Podcast_Ep 54_MORIAH SMALLBONE Becoming a Lead Lady	0.943	0.937	not_laughter
+pf_23761	breath	The Booktopia Books Podcast_John Birmingham on How to Be a Writer	0.927	0.942	laughter
+pf_81497	breath	g1 pdcst_Foodie Call #11 The 2016 Digest	0.942	0.93	laughter
+pf_58668	speech	Spaces Podcast_Evolution of Architecture I	0.906	0.941	laughter
+pf_73853	breath	g1 pdcst_Foodie Call #11 The 2016 Digest	0.941	0.93	laughter
+pf_70708	speech	Free Podcast Day_Home Grown Girls – Apr 7 2018	0.94	0.937	laughter
+pf_80941	speech	Free Podcast Day_Home Grown Girls – Apr 7 2018	0.892	0.939	not_laughter
+pf_56505	breath	Between You & Me Podcast_Ep 54_MORIAH SMALLBONE Becoming a Lead Lady	0.938	0.931	laughter
+pf_75369	speech	Beer Tour Guy Podcast_Episode 12 - Mark Larson - Falling Down Beer Company	0.936	0.937	laughter
+pf_31326	speech	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.937	0.935	laughter
+pf_48743	speech	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.917	0.936	not_laughter
+pf_74113	breath	Design Story_025 Moira Cullen Design is Soft Power	0.935	0.934	laughter
+pf_06727	speech	The Genealogy Happy Hour_Episode 20_Privacy	0.914	0.933	laughter
+pf_57048	breath	Design Story_025 Moira Cullen Design is Soft Power	0.918	0.932	laughter
+pf_75304	breath	Spaces Podcast_Evolution of Architecture I	0.924	0.93	laughter
+pf_23823	speech	Spaces Podcast_Evolution of Architecture I	0.93	0.911	laughter
+pf_01326	breath	The RegenNarration_79. Gondwana Link The most exciting restoration at scale in Australia, with Keith Bradby OAM	0.865	0.929	laughter
+pf_81098	breath	Claire Zorn_'Turns Out I Had More Things To Write Down!'	0.929	0.925	laughter
+pf_10411	speech	Spaces Podcast_Evolution of Architecture I	0.923	0.916	laughter

eval/probs/ami-eval-v1__eat.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f19519a5dbe56f27f18ff8959d4738c3f1c135c39cfdb5b418da4a3b6e69762
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed0.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4422cf6a5f6d0e60ca9680b08f1740caee17bb88c565f4c3692cb01a1aaca8b
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed1.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe3803baa7acc3b653998fb1ac749c6a9965d34cc2dc09670c81d6777f2dd05e
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed10.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:502abd36b110be9748a48e403d7dffb0104ee7231424c56e9938cb788e768e59
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed11.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5733b3ac3529829aee9ca9952a56064c5f24028d8ff42c31315a8ebb806eda4
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed12.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f571cd69f69dcfc22669b89db71c3a0f2f4807a9da82f62321c5ce65156356d5
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed13.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:753b186c8c6cc33780c9d5a5c3f48a04134d0c0e0a5a8ec8beed76080ef16ac0
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed14.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7615eb564c069136840dbaf14a70cf8ccccc239c586a33a826491855c094b2e7
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed2.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:932f33fcb4f453ed3a1e0aad096c36953c75dabeb939e8b0a366975c64ac1f08
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed3.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:781f81b6429da97c3dfd0ea8f113478cc1a17590b4506e402e422abb7e8df7be
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed4.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c911a0099294aee919a9e2e36e1f8e2c4d603343334fd30047413ac4d52b766
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed5.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:152dbddd8313ed29a478eadc2e6622c3724df014fb9f5b6c475169e89e3a280d
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed6.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24cff8270393ce488f3ba66253931af4cff6ecd3394326601556c1981f0296f2
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed7.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f3f8da6e79c6fd93b7bf96d54e805037517d5f882436246540d0c6b0b737350
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed8.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ac5165f5a0359158603e14d39d2cf80ecdcf1f06226d664cbadb8df8eb9502d
+size 12456

eval/probs/ami-eval-v1__laughter-v0.1-seed9.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6130e836221f7ff07841ae53641613bbda29be8cb76a4c9d2717eeed8fb6148
+size 12456

eval/probs/fsd50k-panel__ced-base.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5bddab84987dfc5c47dbb287bdcfa4d3d6ec8b70c05b681c7a79903b9faad8c
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed0.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4d9e8ac86d4a49d9fbb5575b1ed0f019253d40f23b317d03b892910d3e97698
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed1.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dae7d3f6055f89a9c7397a668ccbcc28c5ee3326dee46405151124fde56eb130
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed10.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6e3c4c73b2d811f8c8d630de22873e72ba8228ab4d19217d2f11b3ad8f29ffe
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed11.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48c621d99a0936faebc3965d09e4c6e818241bdf2d99d93ed0d858112dfe33cf
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed12.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b11e6e17ced4798c3ac37a95fc7a83399fa0e7f2e13e45942a33f0a9f104f674
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed13.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:716b59703d3d7fc437a54056fb57452baf4b8440fbc85403a88463139f15e1a7
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed14.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc036d9afca777f5e024153a51f962fee81b4af50cc19900a780e62bb25a74e9
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed2.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95e4e4b212e162fab8600da1f46d71202e1360a743d689d3324db538e2d6537a
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed3.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:985f6a3b53e6513bdbf9fca17476ebe522be7761e895ec3f2a4d61663a9e7424
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed4.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:748496f4327f22cce9925435af9352fda194bf1be7bd2058b9ef1c5d5d0a7491
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed5.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:513cdbff1145c6541454993b1750c6549a49e82255c38a9013cb5a2dd9ac5787
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed6.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0496bac5a1eae49f6a7c5ccc277e31715320cddee37683910bea4f611e3ddb7d
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed7.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a3a56d29a1d0f1be384a0cd8f2675e78dbea8ffa6df19b94fa7940f5f7ae990
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed8.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c3c2b01ad594c6010838991fadaf2978fe49b8949f4e5dd06b9691c18849325
+size 34280

eval/probs/fsd50k-panel__laughter-v0.1-seed9.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:692cf31c4a85e4313a91df97f5232c4477544b5b81abcc98474d34c57a39f8ec
+size 34280

eval/probs/podcastfillers-eval__gillick.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fcdd03193a6b100159bd174ca63b0fc71caad8108a8e3b2dd9490f26d852e565
+size 41720

eval/probs/podcastfillers-eval__laughter-v0.1-seed0.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:690f122faed5b59971263f3fc2a4b42cab9a6ce385ec5547cc5b18f8eed7ac26
+size 41720

eval/probs/podcastfillers-eval__laughter-v0.1-seed1.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf06ca2884932b54590accc5a7e9f63a837a3ceb101dcc566fa770cc5a4d7886
+size 41720

eval/probs/podcastfillers-eval__laughter-v0.1-seed10.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81f49334d0ea69b34b917b58ee2f7a9debe5f577da76a11dacb4ad32417ed4b1
+size 41720

eval/probs/podcastfillers-eval__laughter-v0.1-seed11.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4818408b6dafa09af8b049a24f48a263049a4f30cc0a10e6858d73fbcd217d2a
+size 41720

eval/probs/podcastfillers-eval__laughter-v0.1-seed12.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c3011889f52acb048cec8bd4e7edaae67487cc34581299010dfc640f40a45b6
+size 41720