HashNuke commited on
Commit
a9a2b83
·
verified ·
1 Parent(s): 1994527

Release from source 6ef9558efcd96b3ef4c53f5b15988653581dbd26

Browse files
Files changed (29) hide show
  1. .gitattributes +2 -0
  2. README.md +85 -122
  3. candidates/e5-phrase-run1/LICENSE +21 -0
  4. candidates/e5-phrase-run1/NOTICE +1 -0
  5. candidates/e5-phrase-run1/README.md +108 -0
  6. candidates/e5-phrase-run1/artifacts.sha256.json +22 -0
  7. candidates/e5-phrase-run1/checkpoint/config.json +38 -0
  8. candidates/e5-phrase-run1/checkpoint/model.safetensors +3 -0
  9. candidates/e5-phrase-run1/checkpoint/tokenizer.json +3 -0
  10. candidates/e5-phrase-run1/checkpoint/tokenizer_config.json +17 -0
  11. candidates/e5-phrase-run1/onnx/metadata.json +32 -0
  12. candidates/e5-phrase-run1/onnx/model-fp32.onnx +3 -0
  13. candidates/e5-phrase-run1/onnx/model-int8.onnx +3 -0
  14. candidates/e5-phrase-run1/onnx/tokenizer.json +3 -0
  15. candidates/e5-phrase-run1/provenance.json +170 -0
  16. candidates/e5-phrase-run1/seed-20261115/config.json +38 -0
  17. candidates/e5-phrase-run1/seed-20261115/model.safetensors +3 -0
  18. candidates/e5-phrase-run1/seed-20261115/tokenizer.json +3 -0
  19. candidates/e5-phrase-run1/seed-20261115/tokenizer_config.json +17 -0
  20. candidates/e5-phrase-run1/seed-20261116/config.json +38 -0
  21. candidates/e5-phrase-run1/seed-20261116/model.safetensors +3 -0
  22. candidates/e5-phrase-run1/seed-20261116/tokenizer.json +3 -0
  23. candidates/e5-phrase-run1/seed-20261116/tokenizer_config.json +17 -0
  24. candidates/e5-phrase-run1/seed-20261117/config.json +38 -0
  25. candidates/e5-phrase-run1/seed-20261117/model.safetensors +3 -0
  26. candidates/e5-phrase-run1/seed-20261117/tokenizer.json +3 -0
  27. candidates/e5-phrase-run1/seed-20261117/tokenizer_config.json +17 -0
  28. release.json +50 -28
  29. releases/e5-small-run1.json +37 -0
.gitattributes CHANGED
@@ -1,3 +1,5 @@
1
  *.safetensors filter=lfs diff=lfs merge=lfs -text
2
  /tokenizer.json filter=lfs diff=lfs merge=lfs -text
 
 
3
  *.csv whitespace=cr-at-eol
 
1
  *.safetensors filter=lfs diff=lfs merge=lfs -text
2
  /tokenizer.json filter=lfs diff=lfs merge=lfs -text
3
+ **/tokenizer.json filter=lfs diff=lfs merge=lfs -text
4
+ *.onnx filter=lfs diff=lfs merge=lfs -text
5
  *.csv whitespace=cr-at-eol
README.md CHANGED
@@ -5,142 +5,105 @@ library_name: transformers
5
  pipeline_tag: text-classification
6
  base_model: intfloat/multilingual-e5-small
7
  datasets: [HashNuke/barge-in-detector-data]
8
- tags: [barge-in, backchannel, multilingual, experimental]
9
  ---
10
 
11
- # Barge-in detector
12
 
13
- **Give acknowledgments room. Give interruptions the floor.**
 
 
 
 
14
 
15
- [![GitHub source](https://img.shields.io/badge/GitHub-source-181717?logo=github&logoColor=white)](https://github.com/vxpipe/barge-in-detector)
16
- [![Hugging Face model](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-model-FFD21E)](https://huggingface.co/HashNuke/barge-in-detector)
17
- [![Hugging Face dataset](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-dataset-FFD21E)](https://huggingface.co/datasets/HashNuke/barge-in-detector-data)
18
- [![MIT licence](https://img.shields.io/badge/license-MIT-3DA639)](https://github.com/vxpipe/barge-in-detector/blob/main/LICENSE)
19
 
20
- Voice agents hear callers say “uh-huh”, “yes”, and “okay” while they are still
21
- speaking. Those acknowledgments should let the agent continue. Corrections,
22
- questions, and requests to stop should give the caller the floor.
23
-
24
- Barge-in detector is an open-source, multilingual, text-only project exploring
25
- that distinction with synthetic data and a learned classifier that runs on CPU.
26
-
27
- This repository is the **model release**: weights, tokenizer, configuration,
28
- licence, and provenance. Training, inference scripts, evaluation, benchmarks,
29
- and labnotes are maintained in the [GitHub source](https://github.com/vxpipe/barge-in-detector).
30
-
31
- > **Experimental checkpoint.** The first model misses 25 of 30 short
32
- > interruptions in the reviewed English challenge. Use it for research and
33
- > evaluation; its probabilities are not calibrated.
34
-
35
- ## Run the model
36
 
37
- The CPU example below was tested on Linux with Python 3.14 and uv:
 
 
 
38
 
39
- ```shell
40
- uv venv --python 3.14
41
- uv pip install torch==2.15.0 --index-url https://download.pytorch.org/whl/cpu
42
- uv pip install transformers==5.23.0 huggingface-hub==2.4.0
43
- source .venv/bin/activate
44
- ```
45
 
46
- Run this in Python (`python`) or a notebook using that environment:
47
 
48
  ```python
49
  import torch
50
  from transformers import AutoModelForSequenceClassification, AutoTokenizer
51
 
52
  repo = "HashNuke/barge-in-detector"
53
- tokenizer = AutoTokenizer.from_pretrained(repo)
54
- model = AutoModelForSequenceClassification.from_pretrained(repo).cpu().eval()
55
-
56
- inputs = tokenizer("please stop for a moment", return_tensors="pt", truncation=True, max_length=64)
57
  with torch.inference_mode():
58
- logits = model(**inputs).logits
59
- p_barge_in = logits.softmax(-1)[0, model.config.label2id["barge_in"]].item()
60
-
61
- print({"p_barge_in": p_barge_in, "label": "barge_in" if p_barge_in >= 0.5 else "backchannel"})
62
  ```
63
 
64
- Load once and reuse the model/tokenizer in a server. For a ready-to-run CLI and
65
- training workflow, follow the [GitHub quick start](https://github.com/vxpipe/barge-in-detector#try-the-model-on-your-cpu).
66
- Pin `revision=MODEL_COMMIT` when deploying a specific release.
67
-
68
- ## Input and output
69
-
70
- | Contract | Value |
71
- | --- | --- |
72
- | Input | Caller text spoken over an ongoing agent turn |
73
- | Cleaning | Lowercase, punctuation removed; native scripts may mix with English |
74
- | Budget | First 64 tokenizer tokens |
75
- | Labels | `0 = backchannel`, `1 = barge_in` |
76
- | Output | `p_barge_in` in [0, 1]; current probabilities are uncalibrated |
77
- | Context | No audio, preceding agent line, or language hint is required |
78
-
79
- The runtime handles speech after the agent finishes and owns decision thresholds.
80
- See the [source cleaner contract](https://github.com/vxpipe/barge-in-detector/blob/main/docs/transcript-format.md).
81
- The 0.5 threshold in the example is diagnostic.
82
-
83
- ## Evaluation
84
-
85
- | Evaluation | Accuracy | Missed interruptions | False interruptions |
86
- | --- | --- | --- | --- |
87
- | Synthetic validation (333) | 99.4% | 0.45% | 0.90% |
88
- | Synthetic test (331) | 100% | 0% | 0% |
89
- | Reviewed English challenge (98) | 73.5% | 54.2% | 0% |
90
- | Challenge unseen in any synthetic split (93) | 72.0% | 54.2% | 0% |
91
-
92
- The short-interruption slice is 5/30 correct. The challenge was authored before
93
- inference, then owner-reviewed and reconciled with authorization. Its final
94
- binary labels match the frozen expectations. It includes repeated prefixes and
95
- is now a development/regression set. Five challenge examples appear in synthetic
96
- train/validation. These results are not a representative independent accuracy
97
- estimate. Synthetic splits share criteria; template/semantic leakage remains
98
- unaudited. Human-verified accuracy across all languages is not established.
99
-
100
- ## CPU performance
101
-
102
- Four virtual AMD EPYC Genoa cores, four threads, PyTorch 2.15.0+cpu eager fp32,
103
- batch one, 100 warm-ups and 1,000 measured calls per scenario:
104
-
105
- | Warm text through label | p50 | p95 | p99 |
106
- | --- | --- | --- | --- |
107
- | Typical multilingual test input | 13.1 ms | 19.4 ms | 25.3 ms |
108
- | Full 64-token budget | 26.8 ms | 33.3 ms | 37.9 ms |
109
-
110
- A fresh Python process took **4.22 seconds median** to return its first label
111
- (range 3.85–4.81 seconds over ten runs). That includes imports and local loading;
112
- OS file cache may be warm. Peak warm process RSS was approximately 1,164 MiB.
113
- HTTP, network, queueing, and Vxpipe boundary time are excluded. Reference-CPU
114
- gates are unestablished; ONNX/int8 is unmeasured.
115
-
116
- ## Languages and provenance
117
-
118
- The training/evaluation data spans English (US, India, Singapore), Hindi,
119
- Kannada, Tamil, Telugu, Malayalam, Mandarin, Japanese, Indonesian, Filipino,
120
- French, German, Spanish, and Italian. Results across these 14 languages are
121
- synthetic; the reviewed challenge is English. See the
122
- [dataset card](https://huggingface.co/datasets/HashNuke/barge-in-detector-data) for verification limits.
123
-
124
- The checkpoint fine-tunes [intfloat/multilingual-e5-small](https://huggingface.co/intfloat/multilingual-e5-small)
125
- with a two-label head. First run: 2,640 train / 333 validation / 331 test examples;
126
- five epochs, batch 32, weighted cross-entropy, AdamW, learning rate 3e-5, 10% warm-up,
127
- seed 20261115, bf16 on one RTX 5090. Epoch 5 was selected using validation macro F1
128
- (0.99324). Released weights are fp32. See LICENSE and NOTICE for MIT attribution.
129
-
130
- [release.json](release.json) pins the code and dataset revisions, artifact hashes,
131
- original training-source manifest, and full training metrics. Reports and
132
- labnotes live in the linked GitHub source. This repository contains no training
133
- data or development scripts.
134
-
135
- ## One project, three repositories
136
-
137
- | What you want to do | Start here |
138
- | --- | --- |
139
- | Generate data, train, evaluate, benchmark, or contribute | [GitHub source](https://github.com/vxpipe/barge-in-detector) |
140
- | Download a released model and tokenizer | [Hugging Face model](https://huggingface.co/HashNuke/barge-in-detector) |
141
- | Load a versioned dataset | [Hugging Face dataset](https://huggingface.co/datasets/HashNuke/barge-in-detector-data) |
142
-
143
- GitHub owns all code, tests, docs, labnotes, and milestones, plus a dataset copy.
144
- Hugging Face hosts model and dataset releases. Release metadata links artifacts
145
- to exact source and dataset revisions. The project is available under the
146
- [MIT licence](https://github.com/vxpipe/barge-in-detector/blob/main/LICENSE).
 
5
  pipeline_tag: text-classification
6
  base_model: intfloat/multilingual-e5-small
7
  datasets: [HashNuke/barge-in-detector-data]
8
+ tags: [barge-in, backchannel, multilingual, experimental, onnx]
9
  ---
10
 
11
+ # Multilingual barge-in detector
12
 
13
+ A text classifier for caller speech **over an ongoing agent turn**: backchannel
14
+ (`0`, let the agent continue) or barge-in (`1`, give the caller the floor).
15
+ It consumes the first 64 tokenizer tokens of cleaned caller text, without a
16
+ `query:` prefix, preceding agent line or language hint. Probabilities remain
17
+ uncalibrated; .5 is a diagnostic threshold.
18
 
19
+ The new **E5 phrase-coverage model** is trained on the combined original dataset
20
+ and phrase supplement, across all 14 languages.
 
 
21
 
22
+ | Format | Download |
23
+ | --- | --- |
24
+ | Unquantized fine-tune, FP32 PyTorch | [checkpoint/model.safetensors](candidates/e5-phrase-run1/checkpoint/model.safetensors) |
25
+ | FP32 ONNX | [onnx/model-fp32.onnx](candidates/e5-phrase-run1/onnx/model-fp32.onnx) |
26
+ | Quantized INT8 ONNX | [onnx/model-int8.onnx](candidates/e5-phrase-run1/onnx/model-int8.onnx) |
 
 
 
 
 
 
 
 
 
 
 
27
 
28
+ Tokenizer/configuration accompany the checkpoint. The ONNX directory includes
29
+ its tokenizer and input/output metadata. Quantization applies to matrix
30
+ operations; embeddings remain FP32. See the [candidate card](candidates/e5-phrase-run1/README.md)
31
+ and [verified hashes](candidates/e5-phrase-run1/artifacts.sha256.json).
32
 
33
+ **The root checkpoint remains the original baseline.** Load the subfolder below
34
+ to use the new fine-tune. Original provenance is preserved in
35
+ [releases/e5-small-run1.json](releases/e5-small-run1.json); old weights remain intact.
 
 
 
36
 
37
+ ## Load the new FP32 model on CPU
38
 
39
  ```python
40
  import torch
41
  from transformers import AutoModelForSequenceClassification, AutoTokenizer
42
 
43
  repo = "HashNuke/barge-in-detector"
44
+ subfolder = "candidates/e5-phrase-run1/checkpoint"
45
+ tokenizer = AutoTokenizer.from_pretrained(repo, subfolder=subfolder)
46
+ model = AutoModelForSequenceClassification.from_pretrained(repo, subfolder=subfolder).cpu().eval()
47
+ inputs = tokenizer("please wait", return_tensors="pt", truncation=True, max_length=64)
48
  with torch.inference_mode():
49
+ p_barge_in = model(**inputs).logits.softmax(-1)[0, 1].item()
 
 
 
50
  ```
51
 
52
+ Pin `revision=MODEL_COMMIT` for reproducibility. Clean inputs using the
53
+ [source transcript contract](https://github.com/vxpipe/barge-in-detector/blob/main/docs/transcript-format.md).
54
+ ONNX Runtime CPU inference uses int64 `input_ids` and `attention_mask`; apply
55
+ softmax to the two output logits and read index 1. Serving code lives in the
56
+ [GitHub source](https://github.com/vxpipe/barge-in-detector).
57
+
58
+ ## Training data and evaluation
59
+
60
+ | Split | Original + supplement | Combined |
61
+ | --- | ---: | ---: |
62
+ | Train | 2,640 + 2,278 | 4,918 |
63
+ | Validation | 333 + 582 | 915 |
64
+ | Test | 331 + 577 | 908 |
65
+
66
+ Use dataset configuration `original-plus-phrase-v1` at the exact revision in
67
+ [release.json](release.json). The original dataset and standalone supplement
68
+ are also available. The supplement was generated without a judge; multilingual
69
+ labels remain synthetic.
70
+
71
+ The multilingual E5 encoder and binary head are fully fine-tuned. Three seeds
72
+ use five epochs, batch 32, AdamW 3e-5 and class-weighted cross-entropy. Validation
73
+ macro F1 selects seed 20261117, epoch 4 before heldout evaluation.
74
+
75
+ | Evaluation, threshold .5 | Original E5 | New E5 FP32 | New E5 INT8, batch one |
76
+ | --- | ---: | ---: | ---: |
77
+ | Combined synthetic test | 841/908 | 899/908 | 899/908 |
78
+ | Reviewed English development cases | 72/98 | 93/98 | 94/98 |
79
+ | False interruptions | 0/50 | 0/50 | 0/50 |
80
+ | Missed interruptions | 26/48 | 5/48 | 4/48 |
81
+
82
+ The 98 English cases are reused development examples with pre-existing overlaps
83
+ in the original dataset. They are not an independent estimate of real-world or
84
+ multilingual accuracy. Fresh English/Hindi review and realistic-prior calibration
85
+ remain open. Other languages are synthetic-only.
86
+
87
+ Owner-designated CPU, warmed batch one, four threads, 100 warmups/1,000 calls:
88
+
89
+ | Export | Typical p50 / p95 / p99 | Full 64-token p50 / p95 / p99 |
90
+ | --- | --- | --- |
91
+ | FP32 ONNX | 3.86 / 6.88 / 10.62 ms | 13.24 / 17.75 / 20.82 ms |
92
+ | INT8 ONNX | 2.85 / 4.45 / 5.69 ms | 9.76 / 13.88 / 18.85 ms |
93
+
94
+ INT8 meets the 20 ms model-path p99 target in this measurement. The Vxpipe boundary
95
+ is unmeasured, so the 50 ms whole-path target is not established. INT8 outputs
96
+ can differ from FP32 and with batch composition; these results use batch one.
97
+ This is an experimental release.
98
+
99
+ ## Provenance and source
100
+
101
+ [release.json](release.json) pins the published source and dataset revisions and
102
+ artifact hashes. [Training provenance](candidates/e5-phrase-run1/provenance.json)
103
+ retains the actual source snapshot, base revision and selection.
104
+
105
+ Code, tests, research reports and labnotes live in
106
+ [GitHub](https://github.com/vxpipe/barge-in-detector). Versioned data lives in
107
+ [HashNuke/barge-in-detector-data](https://huggingface.co/datasets/HashNuke/barge-in-detector-data).
108
+ This repository contains model artifacts, cards, licence and provenance only.
109
+ MIT licence; see LICENSE and NOTICE for base-model attribution.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
candidates/e5-phrase-run1/LICENSE ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2026 Akash Manohar
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
candidates/e5-phrase-run1/NOTICE ADDED
@@ -0,0 +1 @@
 
 
1
+ This model is fine-tuned from intfloat/multilingual-e5-small, distributed under the MIT licence. Base model authors and documentation: https://huggingface.co/intfloat/multilingual-e5-small
candidates/e5-phrase-run1/README.md ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: [en, hi, kn, ta, te, ml, zh, ja, id, fil, fr, de, es, it]
3
+ license: mit
4
+ library_name: transformers
5
+ pipeline_tag: text-classification
6
+ base_model: intfloat/multilingual-e5-small
7
+ tags: [barge-in, backchannel, multilingual, experimental, onnx]
8
+ ---
9
+
10
+ # E5 phrase-coverage candidate
11
+
12
+ This candidate fine-tunes multilingual E5 on the original dataset plus a separate
13
+ phrase-coverage supplement. It classifies caller text spoken **over** an agent
14
+ turn as backchannel (`0`) or barge-in (`1`). It uses the first 64 tokenizer tokens,
15
+ without a `query:` prefix, preceding agent line or language hint. Clean inputs
16
+ using the versioned project transcript cleaner.
17
+
18
+ **Experimental release.** See [provenance.json](provenance.json) for the actual training
19
+ snapshot, dataset hashes and selected checkpoint. The original release at the
20
+ repository root is preserved.
21
+
22
+ | Variant | Files | Use |
23
+ | --- | --- | --- |
24
+ | Original unquantized fine-tune | [checkpoint/](checkpoint/) | PyTorch/Transformers, FP32 safetensors |
25
+ | FP32 ONNX | [onnx/model-fp32.onnx](onnx/model-fp32.onnx) | ONNX Runtime CPU |
26
+ | Quantized ONNX | [onnx/model-int8.onnx](onnx/model-int8.onnx) | Dynamic signed INT8 matrix operations; embeddings remain FP32 |
27
+ | Other training seeds | `seed-20261115/`, `seed-20261116/`, `seed-20261117/` | Retained full checkpoints for reproducibility |
28
+
29
+ Tokenizer/configuration accompany the checkpoints. ONNX uses `onnx/tokenizer.json`
30
+ and `onnx/metadata.json`; supply int64 `input_ids` and `attention_mask`, then
31
+ softmax the two output logits and read index 1. The quantized version is itself
32
+ an ONNX export. Its outputs can differ from FP32 and with batch composition.
33
+ The sparse classifier experiments are separate candidates.
34
+
35
+ Load the fine-tuned FP32 model on CPU:
36
+
37
+ ```python
38
+ import torch
39
+ from transformers import AutoModelForSequenceClassification, AutoTokenizer
40
+
41
+ repo = "HashNuke/barge-in-detector"
42
+ subfolder = "candidates/e5-phrase-run1/checkpoint"
43
+ tokenizer = AutoTokenizer.from_pretrained(repo, subfolder=subfolder)
44
+ model = AutoModelForSequenceClassification.from_pretrained(repo, subfolder=subfolder).cpu().eval()
45
+ inputs = tokenizer("please wait", return_tensors="pt", truncation=True, max_length=64)
46
+ with torch.inference_mode():
47
+ p_barge_in = model(**inputs).logits.softmax(-1)[0, 1].item()
48
+ ```
49
+
50
+ Training and evaluation code, research reports and datasets belong in the
51
+ [source repository](https://github.com/vxpipe/barge-in-detector), not this model
52
+ repository. [artifacts.sha256.json](artifacts.sha256.json) records verified model
53
+ file hashes. Licence and base-model attribution are in LICENSE and NOTICE.
54
+
55
+ ## Training and evaluation
56
+
57
+ All 14 languages share fixed 4,918 training / 915 validation / 908 test examples.
58
+ The original dataset is preserved alongside the 3,437-row accepted supplement.
59
+ Base: `intfloat/multilingual-e5-small` at
60
+ `614241f622f53c4eeff9890bdc4f31cfecc418b3`. Full-encoder fine-tuning uses five
61
+ epochs, batch 32, AdamW 3e-5, weight decay .01, 10% warmup, class-weighted
62
+ cross-entropy and BF16 training. The saved checkpoint is FP32.
63
+
64
+ Epoch and seed selection used combined-validation macro F1 before held-out
65
+ analysis: seed 20261117, epoch 4. Each of three seeds trained for 18–22 seconds;
66
+ training, selection and GPU evaluations together took about 70 seconds.
67
+
68
+ | Evaluation, threshold .5 | Original E5 | Selected new E5, FP32 |
69
+ | --- | ---: | ---: |
70
+ | Same combined synthetic test | 841/908 (92.6%) | 899/908 (99.0%) |
71
+ | Reviewed English development cases | 72/98 (73.5%) | 93/98 (94.9%) |
72
+ | False interruptions in reviewed backchannels | 0/50 | 0/50 |
73
+ | Missed reviewed interruptions | 26/48 | 5/48 |
74
+ | Correct short interruptions | 5/30 | 26/30 |
75
+
76
+ Reviewed results across the three new seeds range from 90/98 to 93/98; none
77
+ produces a false interruption on these 50 backchannels. The selected seed was
78
+ chosen on validation, not on its reviewed score.
79
+
80
+ The multilingual test is synthetic. The 98 English examples have been reused
81
+ for development and contain pre-existing overlaps with the original dataset;
82
+ they are not a fresh independent accuracy estimate. Other languages do not yet
83
+ have human-verified acceptance results. Probabilities are uncalibrated under
84
+ realistic traffic priors, and .5 is a diagnostic threshold. Vxpipe whole-path
85
+ latency and fresh human evaluation remain open acceptance gates.
86
+
87
+ ## Local CPU export results
88
+
89
+ Owner-designated AMD EPYC-Genoa CPU, four threads, warmed batch one; 100 warmups
90
+ and 1,000 measurements per scenario. Includes tokenization through probability.
91
+
92
+ | Export | Typical p50 / p95 / p99 | Full 64-token p50 / p95 / p99 |
93
+ | --- | --- | --- |
94
+ | FP32 | 3.86 / 6.88 / 10.62 ms | 13.24 / 17.75 / 20.82 ms |
95
+ | INT8 | 2.85 / 4.45 / 5.69 ms | 9.76 / 13.88 / 18.85 ms |
96
+
97
+ INT8 meets the 20 ms model-path p99 gate in this measurement; FP32 misses it at
98
+ the full input budget. This excludes the Vxpipe boundary, whose 50 ms whole-path
99
+ gate remains unmeasured. Peak process RSS across both loaded formats was about
100
+ 1,173 MiB; it is not an isolated per-model memory measurement. Files are 470.7 MB
101
+ (FP32 ONNX) and 406.8 MB (INT8 ONNX); the large embedding table stays FP32.
102
+
103
+ At serving batch size one, INT8 scores **94/98** on the reviewed set: no false
104
+ interruptions and four missed interruptions, including 27/30 correct short
105
+ requests. It changes one reviewed decision and two synthetic-test decisions
106
+ relative to FP32; synthetic test accuracy remains 899/908. Probability drift can
107
+ be large on individual examples. Batch-32 verification gives different INT8
108
+ outputs, so these reported serving results explicitly use batch one.
candidates/e5-phrase-run1/artifacts.sha256.json ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checkpoint/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
3
+ "checkpoint/model.safetensors": "e83219c637d2c179f87fbf67749dd48fbc1561f569fa741fbab56c224cef9959",
4
+ "checkpoint/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
5
+ "checkpoint/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f",
6
+ "onnx/metadata.json": "5a1a8d27a010860f9395d4c02488f1da2c8bc4ab54cddd7b9d8210123815a0d1",
7
+ "onnx/model-fp32.onnx": "185596eb5d57ffaac87334ca842a96894ff1cbfb3914f4056d73fa3c84f68ceb",
8
+ "onnx/model-int8.onnx": "d8b8f606e5fda9ef9b64c908c222402e3385d9d46f4ff00f2841aa7493845c6b",
9
+ "onnx/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
10
+ "seed-20261115/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
11
+ "seed-20261115/model.safetensors": "c2f087ba8db2b387d269a3ddf8130d42ef954f33ce2c0a2f6fe52eafa9234d80",
12
+ "seed-20261115/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
13
+ "seed-20261115/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f",
14
+ "seed-20261116/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
15
+ "seed-20261116/model.safetensors": "252b4e6d30d797313c0151e9d5e0a6afb13f36d55363946ef037ebf4da595f07",
16
+ "seed-20261116/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
17
+ "seed-20261116/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f",
18
+ "seed-20261117/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
19
+ "seed-20261117/model.safetensors": "e83219c637d2c179f87fbf67749dd48fbc1561f569fa741fbab56c224cef9959",
20
+ "seed-20261117/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
21
+ "seed-20261117/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f"
22
+ }
candidates/e5-phrase-run1/checkpoint/config.json ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_cross_attention": false,
3
+ "architectures": [
4
+ "BertForSequenceClassification"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "bos_token_id": null,
8
+ "classifier_dropout": null,
9
+ "dtype": "float32",
10
+ "eos_token_id": null,
11
+ "hidden_act": "gelu",
12
+ "hidden_dropout_prob": 0.1,
13
+ "hidden_size": 384,
14
+ "id2label": {
15
+ "0": "backchannel",
16
+ "1": "barge_in"
17
+ },
18
+ "initializer_range": 0.02,
19
+ "intermediate_size": 1536,
20
+ "is_decoder": false,
21
+ "label2id": {
22
+ "backchannel": 0,
23
+ "barge_in": 1
24
+ },
25
+ "layer_norm_eps": 1e-12,
26
+ "max_position_embeddings": 512,
27
+ "model_type": "bert",
28
+ "num_attention_heads": 12,
29
+ "num_hidden_layers": 12,
30
+ "pad_token_id": 0,
31
+ "position_embedding_type": "absolute",
32
+ "tie_word_embeddings": true,
33
+ "tokenizer_class": "XLMRobertaTokenizer",
34
+ "transformers_version": "5.23.0",
35
+ "type_vocab_size": 2,
36
+ "use_cache": true,
37
+ "vocab_size": 250037
38
+ }
candidates/e5-phrase-run1/checkpoint/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e83219c637d2c179f87fbf67749dd48fbc1561f569fa741fbab56c224cef9959
3
+ size 470641664
candidates/e5-phrase-run1/checkpoint/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a
3
+ size 17082987
candidates/e5-phrase-run1/checkpoint/tokenizer_config.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": true,
3
+ "backend": "tokenizers",
4
+ "bos_token": "<s>",
5
+ "clean_up_tokenization_spaces": true,
6
+ "cls_token": "<s>",
7
+ "eos_token": "</s>",
8
+ "is_local": true,
9
+ "local_files_only": false,
10
+ "mask_token": "<mask>",
11
+ "model_max_length": 512,
12
+ "pad_token": "<pad>",
13
+ "sep_token": "</s>",
14
+ "sp_model_kwargs": {},
15
+ "tokenizer_class": "XLMRobertaTokenizer",
16
+ "unk_token": "<unk>"
17
+ }
candidates/e5-phrase-run1/onnx/metadata.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "labels": [
4
+ "backchannel",
5
+ "barge_in"
6
+ ],
7
+ "max_length": 64,
8
+ "pad_id": 1,
9
+ "pad_token": "<pad>",
10
+ "model_file": "model-int8.onnx",
11
+ "calibrated": false,
12
+ "threshold": 0.5,
13
+ "quantization": "Dynamic signed int8, per-channel; MatMul/Gemm/Attention. Embeddings remain fp32.",
14
+ "fused_operator_counts": {
15
+ "EmbedLayerNormalization": 1,
16
+ "Attention": 12,
17
+ "MultiHeadAttention": 0,
18
+ "Gelu": 0,
19
+ "FastGelu": 0,
20
+ "BiasGelu": 12,
21
+ "GemmFastGelu": 0,
22
+ "LayerNormalization": 0,
23
+ "SimplifiedLayerNormalization": 0,
24
+ "SkipLayerNormalization": 24,
25
+ "SkipSimplifiedLayerNormalization": 0,
26
+ "RotaryEmbedding": 0,
27
+ "QOrderedAttention": 0,
28
+ "QOrderedGelu": 0,
29
+ "QOrderedLayerNormalization": 0,
30
+ "QOrderedMatMul": 0
31
+ }
32
+ }
candidates/e5-phrase-run1/onnx/model-fp32.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:185596eb5d57ffaac87334ca842a96894ff1cbfb3914f4056d73fa3c84f68ceb
3
+ size 470654766
candidates/e5-phrase-run1/onnx/model-int8.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d8b8f606e5fda9ef9b64c908c222402e3385d9d46f4ff00f2841aa7493845c6b
3
+ size 406787634
candidates/e5-phrase-run1/onnx/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a
3
+ size 17082987
candidates/e5-phrase-run1/provenance.json ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "Experimental release; uncalibrated; fresh human evaluation and Vxpipe whole-path acceptance remain open",
3
+ "base_model": "intfloat/multilingual-e5-small",
4
+ "base_revision": "614241f622f53c4eeff9890bdc4f31cfecc418b3",
5
+ "code_base_commit": "db838c3c48051a18db753a01ad6d01a5171786c9",
6
+ "training_source_uncommitted": true,
7
+ "training_source_sha256": {
8
+ "AGENTS.md": "d1fe86039b9d8050dc313d767b39a90be279434322cafd9409ea915d8c4b2b23",
9
+ "barge_in_detector/__init__.py": "28de6fe1d126dec7b9cd85273f805831b24fb7c443ed2707c47552caede29072",
10
+ "barge_in_detector/__main__.py": "8efaf592230270b84d1ba8513bdb244b104aba73a517f0734197ccbfa18afb8a",
11
+ "barge_in_detector/catalogue.py": "8b59351941bce2a92420ea0b13989a4c002a52d23c6279c10331645fac9796e2",
12
+ "barge_in_detector/cleaner.py": "58bcb1c3f21f66aecb534cbc20e07d5350c3d8d4c92e2084c4943ca9493c3a22",
13
+ "barge_in_detector/cli.py": "cb1aeccf0662f76caee40b3476947e2291cec62d688e5af3438972c7c0d8e4c0",
14
+ "barge_in_detector/generation/__init__.py": "b81aff66375501799977abd59786a5ad9da99f6998febc5fac64a016a6cea3fd",
15
+ "barge_in_detector/generation/config.py": "844d5945b93c44ff04a35f6f161a1ab29a29d39ffacc7af1b268276818d20662",
16
+ "barge_in_detector/generation/llm.py": "540158ae89360fd7697aa357919530cce97f32572c0955307c1986191ae22dca",
17
+ "barge_in_detector/generation/phrase_coverage.py": "350f59c0fe3ff4ee25d279d05e79e4291c03e52d26c4bd62f4d5569fb02fa3e4",
18
+ "barge_in_detector/generation/phrase_runner.py": "612206671e4c68edb37b66ae8dda7eb3fa7f5bb1f48ee3ae07900c35a3be512f",
19
+ "barge_in_detector/generation/prompts.py": "04c3b61b4ef0a7cbd5edcb0438f544b564c602e0d213291dafbcff9fe1f086ce",
20
+ "barge_in_detector/generation/responses.py": "cb5849ddbc90e800bd8c20a17896c0f7cb406b6eacf7126ddc2a1d6468c1ef5d",
21
+ "barge_in_detector/generation/utterance_criteria.py": "78b938e758fa5f8d8d6dedbc48096747a8521eca6812822ef8cccfd70ff62c72",
22
+ "barge_in_detector/generation/utterance_runner.py": "51d383bee865e79242997a6ef1a3ac2add896fb86969689cdbd7a8d26cf734a4",
23
+ "barge_in_detector/generation/utterances.py": "535f7c54e9dff1ed61903713b8c825e98b453932b753da61b46d41479374cd42",
24
+ "barge_in_detector/ids.py": "7a5abdf7ebfb64e5fb1588f887fd951a0df345d71c90d59b445ef2a8c1bf214e",
25
+ "barge_in_detector/releases.py": "e17d8a776d21b3e036c42b01524956fee90ad6b129456b642636b28e142b3cf8",
26
+ "barge_in_detector/review/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
27
+ "barge_in_detector/review/__main__.py": "174a4e77d7534c0136f697395d6b85ad72f862a259f5718a952cef172922fc47",
28
+ "barge_in_detector/review/store.py": "f1147024ddfb910dc4027e74610786bab3b2a5954986ea47e764ae55d1f8b1f0",
29
+ "barge_in_detector/splitting.py": "44e05118111617a5649d94d09e82228864f2f2f5405a15dde974db78d9829a29",
30
+ "barge_in_detector/storage/__init__.py": "4e44293470af3982b64c05c2d17a949185c10b4bd53bf72b5ba6229f51507db4",
31
+ "barge_in_detector/storage/schema.py": "0ccc76ec546a0fe1c414fed1f9f866b0103a55319f04e6d5d3bed12e83862e03",
32
+ "barge_in_detector/storage/store.py": "b3b32e195dc06653702e307d7199cc380b4a74dd687ab860836d491c8166311f",
33
+ "barge_in_detector/supplements.py": "a17039ea5a1bc72ee16b37d5d2fe30ef697dce174b29601193387ff320dfd52c",
34
+ "barge_in_detector/training/__init__.py": "22adb2a7658035c9b7a0a9e8859b30322278830f366c058846a140f92d306a4f",
35
+ "barge_in_detector/training/data.py": "6bd7dbf60fe75d80687f7eb097cdbf03d8a3237197eb07b7f1cfa7dab948e984",
36
+ "barge_in_detector/training/evaluate.py": "eee0aeb17dd13d43cefbad7baa1a6429f1a7aae2c39117e203dbb0515a0bb901",
37
+ "barge_in_detector/training/frozen_e5.py": "23df68d71a479f3548b007457306ae96d0a9d34840d23e64ec9c38b68dc9fc06",
38
+ "barge_in_detector/training/infer.py": "3b7b6a20666c337ced513f40d4d71ee6c0cc0ba69e45f6f6f71e73668eccf4ce",
39
+ "barge_in_detector/training/lexical.py": "835531bd89420f9ada20868d6947edfa2c20aadf63f635e2db5c28516ab88209",
40
+ "barge_in_detector/training/metrics.py": "6909e1cf3e4ac6b763c97ea9a8c26b5ee751a01fffc86c308604d5e437bbfad6",
41
+ "barge_in_detector/training/onnx_export.py": "b6adf895befaf3c6934fe4653cbd17940910d99f9b04d1461945204b055fb6c9",
42
+ "barge_in_detector/training/onnx_infer.py": "abb12f5b0731d5a54358e31460652948874411e6b80306c1f45e295aebc66d13",
43
+ "barge_in_detector/training/train.py": "026de4193f908177612490743510287126f04a6a6b371a3151e3d4cc95a8862f",
44
+ "data/evaluation/english-challenge-v1.csv": "3d46b9e024f4d3bedd69f07603cde934e85ec2272988e90d2acebc30d5e99519",
45
+ "data/evaluation/reviews/english-challenge-v1.json": "ee6e07c3218ab0412d179e73b8e7457d550447a7ae08db2a005f37d794b855fa",
46
+ "data/experiments/original-plus-phrase-v1/splits/test.csv": "9e5315c163b0a4f0151bbf3609013f328a99e89aa0a192ce6f13bddc60bf33a9",
47
+ "data/experiments/original-plus-phrase-v1/splits/train.csv": "8ea73a4c9c22d9ef1b38ee64221c1e322c3b84ae457916080b13cf9c5a7cf751",
48
+ "data/experiments/original-plus-phrase-v1/splits/validation.csv": "2aff187ae2de0fd48d1b2bade5e8b4097a3fc008e9c52a8ff1065fda66816ad7",
49
+ "data/splits/test.csv": "8b630782362b05249b3be94b0213850258422be5ddeb62fcf703817b0bf9e11a",
50
+ "data/splits/train.csv": "aab224bd0526eee915295e6f5c52076e29bb751cd0d664eefb4321843ad22533",
51
+ "data/splits/validation.csv": "c6b4c526fb4b1df96754119e409cedc8c7e85c77659e1fc1c3829da70084edc1",
52
+ "data/supplements/phrase-coverage-v1/splits/test.csv": "b4d8ca14a49b9b799df7608ec4236feea7d0b949ff1cbfd2e35441c24b567ea4",
53
+ "data/supplements/phrase-coverage-v1/splits/train.csv": "4825babe5704fd5b9fef8a8c81c0d05291cdaf0d7eb4469266b582a95bb8ddc3",
54
+ "data/supplements/phrase-coverage-v1/splits/validation.csv": "b8492ad9375c4628872bd2e75ad4cf33f9d3a662d32809c6371a08cd7b6cc134",
55
+ "docs/e5-supplement-finetune.md": "45fd8f2e38cad39978360d0146c4b7c82826b7531da4fc92b3f6d22aa956fe1a",
56
+ "experiments/audit_phrase_coverage.py": "19a6358a495c02b2cc4cc5d442739095d1647f8e133322be7971d256b6b5acf1",
57
+ "experiments/audit_reviewed_errors.py": "e1398c58adcfa83253bbc0ac8e8d18bcf72cf914ca06246558ee9ad7dd4576e2",
58
+ "experiments/benchmark_cpu.py": "f7314faa1dad2ccfbf2040fa11902cee4dcba25bf096ca859febf42bf923037c",
59
+ "experiments/benchmark_cpu_cold.py": "e8ffa6f002d7314e6c52fb2ea1fa0600f1a658caa5f80cf210456bca52310354",
60
+ "experiments/benchmark_frozen_e5.py": "d9a99507c945c418ff8712baf2d773fd9a989dcee934f894033c54c9eb6c7428",
61
+ "experiments/compare_character_baseline.py": "be014c9d82e35003fcfb797cdaebd67398f10772e97032f21b748d1fe6a68669",
62
+ "experiments/compare_frozen_e5.py": "0ee5b50eda9229595e2e24d8cd60ffa4fdd27f4e420149840776cea82e8c8756",
63
+ "experiments/compare_hybrid.py": "17888c41deae4d7daec783b5112237ea110045cbef4a255c6ed12578ae651096",
64
+ "experiments/compare_lexical_features.py": "064c22e06eed92cc3ffcdc03d43d1011711b4286f71430f30ce1f8432ab7f25b",
65
+ "experiments/compare_phrase_supplement.py": "3e47438049b9e1e38937f8025efb913d2132bd277ade8b4eaead75792b0b1527",
66
+ "experiments/evaluate_english_challenge.py": "b924932b58fa218329283353b53507ecd5705b740af2c9a58371253a491107ba",
67
+ "experiments/finetune_e5_supplement.py": "504528b96f8471290403c664c2d5814874c81b8f6288dc5b634bca07155b5a1b",
68
+ "experiments/verify_e5_export.py": "ae5c3e207d7343c87414bb9b1bae49f98334c52b2b6e57bb47f9619386224a20",
69
+ "experiments/verify_lexical.py": "3868e7d88c9f52d18b2e47eb52380c363c54a9377a2388cc31244da17bae2f4f",
70
+ "generation/factors/catalogue.yaml": "1135f3a3bf6000eb1f9d41b1957a348c760e78f96fbe61a8829aa818be0de5b5",
71
+ "generation/factors/constraints.yaml": "422318e1010483e4c5f7eb0de99170eb906b8d21927183a5d6581f13c21ac59c",
72
+ "generation/factors/domains.yaml": "19605dbdd0bb4044c481d74ce5c798534d61dd6c7c46d4edcb2f800c4aa53fc5",
73
+ "generation/factors/intents.yaml": "a3d343addabde711a2395e1ed598f432dd00e547afa962255ae246e17874ae63",
74
+ "generation/factors/languages.yaml": "c5f0a5b753c588782c536108184b12615a54c9b8f4b4cbf0073c0d8aec1a376e",
75
+ "generation/factors/utterances.yaml": "a045546c59d59588a261b03e1f1dc9dc147146f93d3381a813c0920825c435c4",
76
+ "pyproject.toml": "7c15ce2e5e10353e31e645748b567b36023dc582c63a53efa4ee06af2a4f3b29",
77
+ "reports/e5-small-run1/challenge-predictions.csv": "1fb81be4fcab3e20773d0861f95666d1ff3a91f3acd14c2e79a0d3c482225964",
78
+ "tests/fake_llm.py": "2c604f09ff23619225e0c62a57466f300191fd8391777ac7cf05e5556dee4f33",
79
+ "tests/fixtures/cleaner_golden.yaml": "3e013d3775a619a08432fe6c0ced5558934e4dbcf98a4afa651a0bd4bfc1c1db",
80
+ "tests/test_catalogue.py": "9fbfc7bc848e5db38d34ccb08f6bf10bd0908b21cbcc73d70d337c1e0dbae2ca",
81
+ "tests/test_cleaner.py": "048942dafd6e91772d7b17515606cb6c7f44c2225b876542d8598d19116a3d47",
82
+ "tests/test_frozen_e5.py": "ce87bbe996ef858cee479185e4d6caeeb2efd2d082a91ce61485178e2429ec8b",
83
+ "tests/test_gemini_live.py": "64f4ba3a80503ad7e0a43bfbaf7e43fbecbbc5821a41b189c08da5eff390e884",
84
+ "tests/test_ids.py": "3802dc4bd8b75f8764c1d4eafbc882b614096d5d871ec4f5d001990bb3f87d89",
85
+ "tests/test_lexical.py": "729218dc007485add7a033f70537c76b5133e6f43bde7067779ed37cb72b6d54",
86
+ "tests/test_llm.py": "3be9c1710fea8ad5014630f18e569a09bd4f7caed1b01390204c321aa3aa2a65",
87
+ "tests/test_metrics.py": "072656cb74e8750f1a9a52f5d6dadb2ac3250e5f1cd538d6e64f3bd5a46d9567",
88
+ "tests/test_model_data.py": "cc89a589832f6ed66240900321782901a702a95a873ec1c564dcb2e86bc77b3a",
89
+ "tests/test_onnx_infer.py": "a868269fc3411c34153567ef1337a63b823abe6405caf76968aa7c7dcfec748c",
90
+ "tests/test_phrase_coverage.py": "f6d7332187d1d314bc3c34d448897e201493abdd41e8ba61b3c1601dec88ad25",
91
+ "tests/test_releases.py": "8bd068e0be0303211a20912f8fcfabef151fa8d90aeb216d7c0b07c7fbb0665f",
92
+ "tests/test_review.py": "f18cc6a69b88b8657db6f53d99d51e21faee16b1b3d39cb2761304d080941dcc",
93
+ "tests/test_reviewed_audit.py": "fce0d8a64279aba8c164ca68e9b54b9ff125d1e7b1dae6d0512836a43571ce50",
94
+ "tests/test_split.py": "658f548e06fd1d011753741c6d2ef58220990fa6c90ac7df9f6a9a1fb1ffd2be",
95
+ "tests/test_store.py": "f23ab069512838929f7f0f91f8439ceaac4ba38484a6b235750fd7adf672614d",
96
+ "tests/test_supplements.py": "b50f903c398d4d38585dba51c2c56418a37c2c830c7439dcda0c75af1dcc0dc0",
97
+ "tests/test_utterance_criteria.py": "bb076041b038a6b707a6df61ee708408bec89922c011fef8dce0824991691a53",
98
+ "tests/test_utterances.py": "3a48f590d83cddc4db2f57099f5f90bff3b965195f0aeb1adcd663a27bc1a387",
99
+ "uv.lock": "a2d7862608832b6a176ae19f2b39b5fec2e89421ff9bf9d7cd2ff5c2eda20d9e",
100
+ "experiments/run_e5_remote.sh": "700f7346577b54bcc74fbb9f7df91699da562b05cb0ad2723b4e83103341d412"
101
+ },
102
+ "export_source_update": {
103
+ "reason": "Fix fused ONNX tensor type inference; verify all variants; compare unchanged original checkpoint",
104
+ "sha256": {
105
+ "barge_in_detector/training/onnx_export.py": "3668281fca93a5e5f4399df4c99c933100e60aadbea47e42ba74f000d721473a",
106
+ "experiments/verify_e5_export.py": "e12de63f9acf5ba414d6fb9f863f6f24f463eedb5cc1bf792c72dac181c07793",
107
+ "experiments/evaluate_original_e5.py": "ad80d567e030e53eab73f9c5bad44213f3caf96eef12bf61c4e4b33eaeeee130",
108
+ "tests/test_onnx_export.py": "e06e0882caa9f713b0141c942cb56e0b880f0227cca2b9c44c638aa5559fc4f7"
109
+ }
110
+ },
111
+ "dataset_repo": "HashNuke/barge-in-detector-data",
112
+ "dataset_revision": "02114f9cbdeed78c86dcd58ad13011425d010f8e",
113
+ "dataset_files_sha256": {
114
+ "data/experiments/original-plus-phrase-v1/splits/train.csv": "8ea73a4c9c22d9ef1b38ee64221c1e322c3b84ae457916080b13cf9c5a7cf751",
115
+ "data/experiments/original-plus-phrase-v1/splits/validation.csv": "2aff187ae2de0fd48d1b2bade5e8b4097a3fc008e9c52a8ff1065fda66816ad7",
116
+ "data/experiments/original-plus-phrase-v1/splits/test.csv": "9e5315c163b0a4f0151bbf3609013f328a99e89aa0a192ce6f13bddc60bf33a9"
117
+ },
118
+ "selected_seed": 20261117,
119
+ "selected_epoch": 4,
120
+ "selection_rule": "Highest combined-validation macro F1, seed order breaks ties; fixed before test/challenge loading",
121
+ "split_counts": {
122
+ "train": 4918,
123
+ "validation": 915,
124
+ "test": 908
125
+ },
126
+ "labels": [
127
+ "backchannel",
128
+ "barge_in"
129
+ ],
130
+ "max_length": 64,
131
+ "calibrated": false,
132
+ "threshold": 0.5,
133
+ "artifact_sha256": {
134
+ "checkpoint/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
135
+ "checkpoint/model.safetensors": "e83219c637d2c179f87fbf67749dd48fbc1561f569fa741fbab56c224cef9959",
136
+ "checkpoint/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
137
+ "checkpoint/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f",
138
+ "onnx/metadata.json": "5a1a8d27a010860f9395d4c02488f1da2c8bc4ab54cddd7b9d8210123815a0d1",
139
+ "onnx/model-fp32.onnx": "185596eb5d57ffaac87334ca842a96894ff1cbfb3914f4056d73fa3c84f68ceb",
140
+ "onnx/model-int8.onnx": "d8b8f606e5fda9ef9b64c908c222402e3385d9d46f4ff00f2841aa7493845c6b",
141
+ "onnx/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
142
+ "seed-20261115/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
143
+ "seed-20261115/model.safetensors": "c2f087ba8db2b387d269a3ddf8130d42ef954f33ce2c0a2f6fe52eafa9234d80",
144
+ "seed-20261115/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
145
+ "seed-20261115/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f",
146
+ "seed-20261116/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
147
+ "seed-20261116/model.safetensors": "252b4e6d30d797313c0151e9d5e0a6afb13f36d55363946ef037ebf4da595f07",
148
+ "seed-20261116/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
149
+ "seed-20261116/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f",
150
+ "seed-20261117/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
151
+ "seed-20261117/model.safetensors": "e83219c637d2c179f87fbf67749dd48fbc1561f569fa741fbab56c224cef9959",
152
+ "seed-20261117/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
153
+ "seed-20261117/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f"
154
+ },
155
+ "verified_local_formats": [
156
+ "PyTorch safetensors",
157
+ "FP32 ONNX",
158
+ "dynamic INT8 ONNX"
159
+ ],
160
+ "limitations": [
161
+ "Synthetic multilingual test; English review reused for development",
162
+ "INT8 decisions and probabilities can differ from FP32 and with batch composition",
163
+ "No realistic-prior calibration or Vxpipe whole-path verification"
164
+ ],
165
+ "code_revision": "6ef9558efcd96b3ef4c53f5b15988653581dbd26",
166
+ "source_repo": "https://github.com/vxpipe/barge-in-detector",
167
+ "dataset_config": "original-plus-phrase-v1",
168
+ "training_snapshot_note": "Training began from an uncommitted snapshot based on code_base_commit. Exact per-file training hashes are retained; code_revision identifies its published source and export fixes.",
169
+ "published_training_script_matches_snapshot": true
170
+ }
candidates/e5-phrase-run1/seed-20261115/config.json ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_cross_attention": false,
3
+ "architectures": [
4
+ "BertForSequenceClassification"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "bos_token_id": null,
8
+ "classifier_dropout": null,
9
+ "dtype": "float32",
10
+ "eos_token_id": null,
11
+ "hidden_act": "gelu",
12
+ "hidden_dropout_prob": 0.1,
13
+ "hidden_size": 384,
14
+ "id2label": {
15
+ "0": "backchannel",
16
+ "1": "barge_in"
17
+ },
18
+ "initializer_range": 0.02,
19
+ "intermediate_size": 1536,
20
+ "is_decoder": false,
21
+ "label2id": {
22
+ "backchannel": 0,
23
+ "barge_in": 1
24
+ },
25
+ "layer_norm_eps": 1e-12,
26
+ "max_position_embeddings": 512,
27
+ "model_type": "bert",
28
+ "num_attention_heads": 12,
29
+ "num_hidden_layers": 12,
30
+ "pad_token_id": 0,
31
+ "position_embedding_type": "absolute",
32
+ "tie_word_embeddings": true,
33
+ "tokenizer_class": "XLMRobertaTokenizer",
34
+ "transformers_version": "5.23.0",
35
+ "type_vocab_size": 2,
36
+ "use_cache": true,
37
+ "vocab_size": 250037
38
+ }
candidates/e5-phrase-run1/seed-20261115/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c2f087ba8db2b387d269a3ddf8130d42ef954f33ce2c0a2f6fe52eafa9234d80
3
+ size 470641664
candidates/e5-phrase-run1/seed-20261115/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a
3
+ size 17082987
candidates/e5-phrase-run1/seed-20261115/tokenizer_config.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": true,
3
+ "backend": "tokenizers",
4
+ "bos_token": "<s>",
5
+ "clean_up_tokenization_spaces": true,
6
+ "cls_token": "<s>",
7
+ "eos_token": "</s>",
8
+ "is_local": true,
9
+ "local_files_only": false,
10
+ "mask_token": "<mask>",
11
+ "model_max_length": 512,
12
+ "pad_token": "<pad>",
13
+ "sep_token": "</s>",
14
+ "sp_model_kwargs": {},
15
+ "tokenizer_class": "XLMRobertaTokenizer",
16
+ "unk_token": "<unk>"
17
+ }
candidates/e5-phrase-run1/seed-20261116/config.json ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_cross_attention": false,
3
+ "architectures": [
4
+ "BertForSequenceClassification"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "bos_token_id": null,
8
+ "classifier_dropout": null,
9
+ "dtype": "float32",
10
+ "eos_token_id": null,
11
+ "hidden_act": "gelu",
12
+ "hidden_dropout_prob": 0.1,
13
+ "hidden_size": 384,
14
+ "id2label": {
15
+ "0": "backchannel",
16
+ "1": "barge_in"
17
+ },
18
+ "initializer_range": 0.02,
19
+ "intermediate_size": 1536,
20
+ "is_decoder": false,
21
+ "label2id": {
22
+ "backchannel": 0,
23
+ "barge_in": 1
24
+ },
25
+ "layer_norm_eps": 1e-12,
26
+ "max_position_embeddings": 512,
27
+ "model_type": "bert",
28
+ "num_attention_heads": 12,
29
+ "num_hidden_layers": 12,
30
+ "pad_token_id": 0,
31
+ "position_embedding_type": "absolute",
32
+ "tie_word_embeddings": true,
33
+ "tokenizer_class": "XLMRobertaTokenizer",
34
+ "transformers_version": "5.23.0",
35
+ "type_vocab_size": 2,
36
+ "use_cache": true,
37
+ "vocab_size": 250037
38
+ }
candidates/e5-phrase-run1/seed-20261116/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:252b4e6d30d797313c0151e9d5e0a6afb13f36d55363946ef037ebf4da595f07
3
+ size 470641664
candidates/e5-phrase-run1/seed-20261116/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a
3
+ size 17082987
candidates/e5-phrase-run1/seed-20261116/tokenizer_config.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": true,
3
+ "backend": "tokenizers",
4
+ "bos_token": "<s>",
5
+ "clean_up_tokenization_spaces": true,
6
+ "cls_token": "<s>",
7
+ "eos_token": "</s>",
8
+ "is_local": true,
9
+ "local_files_only": false,
10
+ "mask_token": "<mask>",
11
+ "model_max_length": 512,
12
+ "pad_token": "<pad>",
13
+ "sep_token": "</s>",
14
+ "sp_model_kwargs": {},
15
+ "tokenizer_class": "XLMRobertaTokenizer",
16
+ "unk_token": "<unk>"
17
+ }
candidates/e5-phrase-run1/seed-20261117/config.json ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_cross_attention": false,
3
+ "architectures": [
4
+ "BertForSequenceClassification"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "bos_token_id": null,
8
+ "classifier_dropout": null,
9
+ "dtype": "float32",
10
+ "eos_token_id": null,
11
+ "hidden_act": "gelu",
12
+ "hidden_dropout_prob": 0.1,
13
+ "hidden_size": 384,
14
+ "id2label": {
15
+ "0": "backchannel",
16
+ "1": "barge_in"
17
+ },
18
+ "initializer_range": 0.02,
19
+ "intermediate_size": 1536,
20
+ "is_decoder": false,
21
+ "label2id": {
22
+ "backchannel": 0,
23
+ "barge_in": 1
24
+ },
25
+ "layer_norm_eps": 1e-12,
26
+ "max_position_embeddings": 512,
27
+ "model_type": "bert",
28
+ "num_attention_heads": 12,
29
+ "num_hidden_layers": 12,
30
+ "pad_token_id": 0,
31
+ "position_embedding_type": "absolute",
32
+ "tie_word_embeddings": true,
33
+ "tokenizer_class": "XLMRobertaTokenizer",
34
+ "transformers_version": "5.23.0",
35
+ "type_vocab_size": 2,
36
+ "use_cache": true,
37
+ "vocab_size": 250037
38
+ }
candidates/e5-phrase-run1/seed-20261117/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e83219c637d2c179f87fbf67749dd48fbc1561f569fa741fbab56c224cef9959
3
+ size 470641664
candidates/e5-phrase-run1/seed-20261117/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a
3
+ size 17082987
candidates/e5-phrase-run1/seed-20261117/tokenizer_config.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": true,
3
+ "backend": "tokenizers",
4
+ "bos_token": "<s>",
5
+ "clean_up_tokenization_spaces": true,
6
+ "cls_token": "<s>",
7
+ "eos_token": "</s>",
8
+ "is_local": true,
9
+ "local_files_only": false,
10
+ "mask_token": "<mask>",
11
+ "model_max_length": 512,
12
+ "pad_token": "<pad>",
13
+ "sep_token": "</s>",
14
+ "sp_model_kwargs": {},
15
+ "tokenizer_class": "XLMRobertaTokenizer",
16
+ "unk_token": "<unk>"
17
+ }
release.json CHANGED
@@ -1,37 +1,59 @@
1
  {
 
2
  "model_id": "HashNuke/barge-in-detector",
3
- "status": "experimental, uncalibrated, short-interruption failures",
 
 
 
4
  "base_model": "intfloat/multilingual-e5-small",
 
5
  "dataset_repo": "HashNuke/barge-in-detector-data",
6
- "dataset_revision": "9e914ce8b50aa12ebd3fe212bb183e09a1dee370",
 
7
  "dataset_files_sha256": {
8
- "data/export/utterances.csv": "556a2820e4a24ac13bab44321fb32b464c5fb225750d41a5a10ebcde7cdbab8a",
9
- "data/splits/train.csv": "aab224bd0526eee915295e6f5c52076e29bb751cd0d664eefb4321843ad22533",
10
- "data/splits/validation.csv": "c6b4c526fb4b1df96754119e409cedc8c7e85c77659e1fc1c3829da70084edc1",
11
- "data/splits/test.csv": "8b630782362b05249b3be94b0213850258422be5ddeb62fcf703817b0bf9e11a",
12
- "data/evaluation/english-challenge-v1.csv": "3d46b9e024f4d3bedd69f07603cde934e85ec2272988e90d2acebc30d5e99519",
13
- "data/evaluation/reviews/english-challenge-v1.json": "ee6e07c3218ab0412d179e73b8e7457d550447a7ae08db2a005f37d794b855fa"
14
- },
15
- "artifacts_sha256": {
16
- "model.safetensors": "bddb65c08f6260cfc9d320e65b4229229ec9f3ca95204aed98ab5ae45ff758d3",
17
- "config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
18
- "tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
19
- "tokenizer_config.json": "33b75086cc783ee5d78cab4dae0762d9cab1c40a8e0feb6e65d60c8d69a6afc1"
20
  },
21
- "original_training_source": {
22
- "git_base_commit": "5ce0ed1283401f8c73bafcf5a7dbd2101c4bbee6",
23
- "overlay_manifest": "https://github.com/vxpipe/barge-in-detector/blob/836de6924c3da68cd871e6e08da34291e9778432/reports/e5-small-run1/source-manifest.json"
 
24
  },
25
- "labels": {
26
- "0": "backchannel",
27
- "1": "barge_in"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
28
  },
29
- "max_length": 64,
30
- "default_diagnostic_threshold": 0.5,
31
- "selected_epoch": 5,
32
- "training_metrics": "https://github.com/vxpipe/barge-in-detector/blob/836de6924c3da68cd871e6e08da34291e9778432/reports/e5-small-run1/training-metrics.json",
33
- "source_repo": "https://github.com/vxpipe/barge-in-detector",
34
- "code_revision": "836de6924c3da68cd871e6e08da34291e9778432",
35
- "first_publication_source_commit": "1f92917a59908e320c3a3489c3e5de413a3e4672",
36
- "ownership": "Release artifacts only. All development code, tests, reports and labnotes are in the GitHub source repository."
37
  }
 
1
  {
2
+ "release_id": "e5-phrase-run1-2026-11-15",
3
  "model_id": "HashNuke/barge-in-detector",
4
+ "source_repo": "https://github.com/vxpipe/barge-in-detector",
5
+ "code_revision": "6ef9558efcd96b3ef4c53f5b15988653581dbd26",
6
+ "status": "Experimental release; uncalibrated; fresh human evaluation and Vxpipe whole-path acceptance remain open",
7
+ "model_subdirectory": "candidates/e5-phrase-run1/checkpoint",
8
  "base_model": "intfloat/multilingual-e5-small",
9
+ "base_revision": "614241f622f53c4eeff9890bdc4f31cfecc418b3",
10
  "dataset_repo": "HashNuke/barge-in-detector-data",
11
+ "dataset_revision": "02114f9cbdeed78c86dcd58ad13011425d010f8e",
12
+ "dataset_config": "original-plus-phrase-v1",
13
  "dataset_files_sha256": {
14
+ "data/experiments/original-plus-phrase-v1/splits/train.csv": "8ea73a4c9c22d9ef1b38ee64221c1e322c3b84ae457916080b13cf9c5a7cf751",
15
+ "data/experiments/original-plus-phrase-v1/splits/validation.csv": "2aff187ae2de0fd48d1b2bade5e8b4097a3fc008e9c52a8ff1065fda66816ad7",
16
+ "data/experiments/original-plus-phrase-v1/splits/test.csv": "9e5315c163b0a4f0151bbf3609013f328a99e89aa0a192ce6f13bddc60bf33a9"
 
 
 
 
 
 
 
 
 
17
  },
18
+ "split_counts": {
19
+ "train": 4918,
20
+ "validation": 915,
21
+ "test": 908
22
  },
23
+ "selected_seed": 20261117,
24
+ "selected_epoch": 4,
25
+ "selection_rule": "Highest combined-validation macro F1, seed order breaks ties; fixed before test/challenge loading",
26
+ "artifacts_sha256": {
27
+ "candidates/e5-phrase-run1/LICENSE": "50872f75ca036fc8e1a76621f7b997a6c669d80a5976de8b2bdff7570f3f966f",
28
+ "candidates/e5-phrase-run1/NOTICE": "0103913575e359c247d367d0a0197d1d7f3165ea83ba9ffa0e5d2a584fb695c9",
29
+ "candidates/e5-phrase-run1/README.md": "94b875dfb1f485b63a27ad439245292c4309956161c062d4c3ee59a61ce6c733",
30
+ "candidates/e5-phrase-run1/artifacts.sha256.json": "58a94b54dc8ca36c7e436ff259efd85f4c7d9b7dd5ba32a6ed61f446e5981519",
31
+ "candidates/e5-phrase-run1/checkpoint/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
32
+ "candidates/e5-phrase-run1/checkpoint/model.safetensors": "e83219c637d2c179f87fbf67749dd48fbc1561f569fa741fbab56c224cef9959",
33
+ "candidates/e5-phrase-run1/checkpoint/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
34
+ "candidates/e5-phrase-run1/checkpoint/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f",
35
+ "candidates/e5-phrase-run1/onnx/metadata.json": "5a1a8d27a010860f9395d4c02488f1da2c8bc4ab54cddd7b9d8210123815a0d1",
36
+ "candidates/e5-phrase-run1/onnx/model-fp32.onnx": "185596eb5d57ffaac87334ca842a96894ff1cbfb3914f4056d73fa3c84f68ceb",
37
+ "candidates/e5-phrase-run1/onnx/model-int8.onnx": "d8b8f606e5fda9ef9b64c908c222402e3385d9d46f4ff00f2841aa7493845c6b",
38
+ "candidates/e5-phrase-run1/onnx/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
39
+ "candidates/e5-phrase-run1/provenance.json": "d1bf86951b685b20961d92924adb4d356aea0ba4da484fc7202fd658fa5c6492",
40
+ "candidates/e5-phrase-run1/seed-20261115/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
41
+ "candidates/e5-phrase-run1/seed-20261115/model.safetensors": "c2f087ba8db2b387d269a3ddf8130d42ef954f33ce2c0a2f6fe52eafa9234d80",
42
+ "candidates/e5-phrase-run1/seed-20261115/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
43
+ "candidates/e5-phrase-run1/seed-20261115/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f",
44
+ "candidates/e5-phrase-run1/seed-20261116/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
45
+ "candidates/e5-phrase-run1/seed-20261116/model.safetensors": "252b4e6d30d797313c0151e9d5e0a6afb13f36d55363946ef037ebf4da595f07",
46
+ "candidates/e5-phrase-run1/seed-20261116/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
47
+ "candidates/e5-phrase-run1/seed-20261116/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f",
48
+ "candidates/e5-phrase-run1/seed-20261117/config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
49
+ "candidates/e5-phrase-run1/seed-20261117/model.safetensors": "e83219c637d2c179f87fbf67749dd48fbc1561f569fa741fbab56c224cef9959",
50
+ "candidates/e5-phrase-run1/seed-20261117/tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
51
+ "candidates/e5-phrase-run1/seed-20261117/tokenizer_config.json": "13660843154d3fde0f9954ad198b5a8bbf0d21cd4fc83aa4f37678d224c8481f",
52
+ ".gitattributes": "44402d83165e9763dc772d363f5cc6ef4265f71b62ef71769d20ab4dac1d9609",
53
+ "releases/e5-small-run1.json": "4605d2e1cac8869d35de2049f4aeab0f032b8e5e1acde2015e7373a20c4eb66e"
54
  },
55
+ "evaluation_source": "docs/e5-phrase-results.md",
56
+ "original_baseline_metadata": "releases/e5-small-run1.json",
57
+ "training_snapshot_manifest": "reports/e5-phrase-run1/source-manifest.json",
58
+ "export_update_manifest": "reports/e5-phrase-run1/export-source-update.json"
 
 
 
 
59
  }
releases/e5-small-run1.json ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_id": "HashNuke/barge-in-detector",
3
+ "status": "experimental, uncalibrated, short-interruption failures",
4
+ "base_model": "intfloat/multilingual-e5-small",
5
+ "dataset_repo": "HashNuke/barge-in-detector-data",
6
+ "dataset_revision": "9e914ce8b50aa12ebd3fe212bb183e09a1dee370",
7
+ "dataset_files_sha256": {
8
+ "data/export/utterances.csv": "556a2820e4a24ac13bab44321fb32b464c5fb225750d41a5a10ebcde7cdbab8a",
9
+ "data/splits/train.csv": "aab224bd0526eee915295e6f5c52076e29bb751cd0d664eefb4321843ad22533",
10
+ "data/splits/validation.csv": "c6b4c526fb4b1df96754119e409cedc8c7e85c77659e1fc1c3829da70084edc1",
11
+ "data/splits/test.csv": "8b630782362b05249b3be94b0213850258422be5ddeb62fcf703817b0bf9e11a",
12
+ "data/evaluation/english-challenge-v1.csv": "3d46b9e024f4d3bedd69f07603cde934e85ec2272988e90d2acebc30d5e99519",
13
+ "data/evaluation/reviews/english-challenge-v1.json": "ee6e07c3218ab0412d179e73b8e7457d550447a7ae08db2a005f37d794b855fa"
14
+ },
15
+ "artifacts_sha256": {
16
+ "model.safetensors": "bddb65c08f6260cfc9d320e65b4229229ec9f3ca95204aed98ab5ae45ff758d3",
17
+ "config.json": "e37f926b37b69e1ae223add18a9b0e993d6000cfe87c44f2dd42ce157654d93b",
18
+ "tokenizer.json": "255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a",
19
+ "tokenizer_config.json": "33b75086cc783ee5d78cab4dae0762d9cab1c40a8e0feb6e65d60c8d69a6afc1"
20
+ },
21
+ "original_training_source": {
22
+ "git_base_commit": "5ce0ed1283401f8c73bafcf5a7dbd2101c4bbee6",
23
+ "overlay_manifest": "https://github.com/vxpipe/barge-in-detector/blob/836de6924c3da68cd871e6e08da34291e9778432/reports/e5-small-run1/source-manifest.json"
24
+ },
25
+ "labels": {
26
+ "0": "backchannel",
27
+ "1": "barge_in"
28
+ },
29
+ "max_length": 64,
30
+ "default_diagnostic_threshold": 0.5,
31
+ "selected_epoch": 5,
32
+ "training_metrics": "https://github.com/vxpipe/barge-in-detector/blob/836de6924c3da68cd871e6e08da34291e9778432/reports/e5-small-run1/training-metrics.json",
33
+ "source_repo": "https://github.com/vxpipe/barge-in-detector",
34
+ "code_revision": "836de6924c3da68cd871e6e08da34291e9778432",
35
+ "first_publication_source_commit": "1f92917a59908e320c3a3489c3e5de413a3e4672",
36
+ "ownership": "Release artifacts only. All development code, tests, reports and labnotes are in the GitHub source repository."
37
+ }