Datasets:
AD4Edu v21 (2026-09-12): UltraChat/UltraFeedback shapes, ADEduEval C-mech ratings, train_gen; SFT v10, dpo_train_v20, human_val_400
Browse files- README.md +241 -61
- data/test_human-00000-of-00001.parquet +2 -2
- data/train_prefs-00000-of-00001.parquet +2 -2
README.md
CHANGED
|
@@ -5,22 +5,78 @@ license: other
|
|
| 5 |
license_name: research-only-video-derived
|
| 6 |
task_categories:
|
| 7 |
- text-generation
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
pretty_name: AD4Edu Preferences
|
| 9 |
configs:
|
| 10 |
- config_name: default
|
| 11 |
data_files:
|
| 12 |
- split: train_prefs
|
| 13 |
path: data/train_prefs-*
|
|
|
|
|
|
|
| 14 |
- split: test_human
|
| 15 |
path: data/test_human-*
|
| 16 |
dataset_info:
|
|
|
|
| 17 |
features:
|
|
|
|
|
|
|
| 18 |
- name: prompt
|
| 19 |
dtype: string
|
| 20 |
-
- name:
|
| 21 |
dtype: string
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
- name: rejected
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
dtype: string
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
- name: pair_id
|
| 25 |
dtype: string
|
| 26 |
- name: moment_id
|
|
@@ -33,30 +89,72 @@ dataset_info:
|
|
| 33 |
dtype: string
|
| 34 |
- name: label_source
|
| 35 |
dtype: string
|
| 36 |
-
- name: source
|
| 37 |
-
dtype: string
|
| 38 |
- name: cross_family
|
| 39 |
dtype: bool
|
|
|
|
|
|
|
| 40 |
splits:
|
| 41 |
- name: train_prefs
|
| 42 |
-
num_bytes:
|
| 43 |
num_examples: 3540
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
# Dataset Card for AD4Edu Preferences
|
| 52 |
-
|
| 53 |
-
<!-- test_human features: kept distinct from train_prefs on purpose
|
| 54 |
- name: prompt
|
| 55 |
dtype: string
|
| 56 |
-
- name:
|
| 57 |
dtype: string
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
- name: option_b
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
dtype: string
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
- name: pair_id
|
| 61 |
dtype: string
|
| 62 |
- name: moment_id
|
|
@@ -69,82 +167,164 @@ dataset_info:
|
|
| 69 |
dtype: string
|
| 70 |
- name: stratum
|
| 71 |
dtype: string
|
|
|
|
|
|
|
|
|
|
|
|
|
| 72 |
- name: labels_pending
|
| 73 |
dtype: bool
|
| 74 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
|
| 76 |
-
#
|
| 77 |
|
| 78 |
Preference pairs over **audio descriptions (AD) of slide-based lecture videos**, for blind and
|
| 79 |
-
low-vision students. Each pair is two candidate ADs for the same lecture *moment*; the task is to
|
| 80 |
-
say which better serves a BLV listener under the project's
|
|
|
|
|
|
|
| 81 |
|
| 82 |
**PRIVATE, and derived from copyrighted lecture video.** Do not redistribute. Keyframe images are
|
| 83 |
-
not included here
|
| 84 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
|
| 86 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 87 |
|
| 88 |
-
|
|
|
|
| 89 |
|
| 90 |
```python
|
| 91 |
from datasets import load_dataset
|
| 92 |
-
|
|
|
|
| 93 |
```
|
| 94 |
|
| 95 |
-
##
|
| 96 |
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
|
|
|
| 101 |
|
| 102 |
-
|
| 103 |
-
The labels are not missing, they are *pending*: these are the pairs going to two blind/low-vision
|
| 104 |
-
raters, and putting a machine label on them would substitute a guess for the measurement the split
|
| 105 |
-
exists to make. If you need labelled preference data, use `train_prefs`.
|
| 106 |
|
| 107 |
-
|
| 108 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 109 |
|
| 110 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 111 |
|
| 112 |
-
|
| 113 |
|
| 114 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
|---|---|---|
|
| 116 |
| `construction` | 1039 | one axis perturbed deliberately; direction true by construction, no judge involved |
|
| 117 |
| `judge` | 2501 | Gemini judged a 16 s clip in **both** presentation orders and agreed with itself |
|
| 118 |
|
| 119 |
-
The judge was measured before use
|
| 120 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
-
|
| 123 |
-
presentation orders are mandatory: a single-order run assigns a confident winner to every one.
|
| 124 |
-
* **Longer side wins 56.9%** [54.6, 59.2] on cross-family decided pairs — **+6.9 pp**. Small but
|
| 125 |
-
real: the interval excludes 50%, and it will bias any length-sensitive comparison these labels
|
| 126 |
-
feed. It is nothing like the 77.6% that compromised an earlier release.
|
| 127 |
-
* Self-preference **not** borne out: the judge's own family wins 54.9%, Claude's 67.0%.
|
| 128 |
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
|
|
|
| 135 |
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 139 |
|
| 140 |
-
|
| 141 |
-
|
|
|
|
| 142 |
|
| 143 |
-
##
|
| 144 |
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
|
|
|
|
|
|
| 148 |
|
| 149 |
## Citation
|
| 150 |
|
|
|
|
| 5 |
license_name: research-only-video-derived
|
| 6 |
task_categories:
|
| 7 |
- text-generation
|
| 8 |
+
tags:
|
| 9 |
+
- dpo
|
| 10 |
+
- preference
|
| 11 |
+
- audio-description
|
| 12 |
+
- accessibility
|
| 13 |
pretty_name: AD4Edu Preferences
|
| 14 |
configs:
|
| 15 |
- config_name: default
|
| 16 |
data_files:
|
| 17 |
- split: train_prefs
|
| 18 |
path: data/train_prefs-*
|
| 19 |
+
- config_name: test_human
|
| 20 |
+
data_files:
|
| 21 |
- split: test_human
|
| 22 |
path: data/test_human-*
|
| 23 |
dataset_info:
|
| 24 |
+
- config_name: default
|
| 25 |
features:
|
| 26 |
+
- name: source
|
| 27 |
+
dtype: string
|
| 28 |
- name: prompt
|
| 29 |
dtype: string
|
| 30 |
+
- name: prompt_id
|
| 31 |
dtype: string
|
| 32 |
+
- name: chosen
|
| 33 |
+
list:
|
| 34 |
+
- name: content
|
| 35 |
+
dtype: string
|
| 36 |
+
- name: role
|
| 37 |
+
dtype: string
|
| 38 |
- name: rejected
|
| 39 |
+
list:
|
| 40 |
+
- name: content
|
| 41 |
+
dtype: string
|
| 42 |
+
- name: role
|
| 43 |
+
dtype: string
|
| 44 |
+
- name: chosen-rating
|
| 45 |
+
dtype: float64
|
| 46 |
+
- name: chosen-model
|
| 47 |
dtype: string
|
| 48 |
+
- name: rejected-rating
|
| 49 |
+
dtype: float64
|
| 50 |
+
- name: rejected-model
|
| 51 |
+
dtype: string
|
| 52 |
+
- name: chosen-components
|
| 53 |
+
struct:
|
| 54 |
+
- name: style
|
| 55 |
+
dtype: float64
|
| 56 |
+
- name: terminology
|
| 57 |
+
dtype: float64
|
| 58 |
+
- name: length
|
| 59 |
+
dtype: float64
|
| 60 |
+
- name: deixis
|
| 61 |
+
dtype: float64
|
| 62 |
+
- name: faithfulness
|
| 63 |
+
dtype: float64
|
| 64 |
+
- name: non_redundancy
|
| 65 |
+
dtype: float64
|
| 66 |
+
- name: rejected-components
|
| 67 |
+
struct:
|
| 68 |
+
- name: style
|
| 69 |
+
dtype: float64
|
| 70 |
+
- name: terminology
|
| 71 |
+
dtype: float64
|
| 72 |
+
- name: length
|
| 73 |
+
dtype: float64
|
| 74 |
+
- name: deixis
|
| 75 |
+
dtype: float64
|
| 76 |
+
- name: faithfulness
|
| 77 |
+
dtype: float64
|
| 78 |
+
- name: non_redundancy
|
| 79 |
+
dtype: float64
|
| 80 |
- name: pair_id
|
| 81 |
dtype: string
|
| 82 |
- name: moment_id
|
|
|
|
| 89 |
dtype: string
|
| 90 |
- name: label_source
|
| 91 |
dtype: string
|
|
|
|
|
|
|
| 92 |
- name: cross_family
|
| 93 |
dtype: bool
|
| 94 |
+
- name: what_differs
|
| 95 |
+
dtype: string
|
| 96 |
splits:
|
| 97 |
- name: train_prefs
|
| 98 |
+
num_bytes: 12415042
|
| 99 |
num_examples: 3540
|
| 100 |
+
download_size: 1605312
|
| 101 |
+
dataset_size: 12415042
|
| 102 |
+
- config_name: test_human
|
| 103 |
+
features:
|
| 104 |
+
- name: source
|
| 105 |
+
dtype: string
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
- name: prompt
|
| 107 |
dtype: string
|
| 108 |
+
- name: prompt_id
|
| 109 |
dtype: string
|
| 110 |
+
- name: option_a
|
| 111 |
+
list:
|
| 112 |
+
- name: content
|
| 113 |
+
dtype: string
|
| 114 |
+
- name: role
|
| 115 |
+
dtype: string
|
| 116 |
- name: option_b
|
| 117 |
+
list:
|
| 118 |
+
- name: content
|
| 119 |
+
dtype: string
|
| 120 |
+
- name: role
|
| 121 |
+
dtype: string
|
| 122 |
+
- name: option_a-rating
|
| 123 |
+
dtype: float64
|
| 124 |
+
- name: option_a-model
|
| 125 |
dtype: string
|
| 126 |
+
- name: option_b-rating
|
| 127 |
+
dtype: float64
|
| 128 |
+
- name: option_b-model
|
| 129 |
+
dtype: string
|
| 130 |
+
- name: option_a-components
|
| 131 |
+
struct:
|
| 132 |
+
- name: style
|
| 133 |
+
dtype: float64
|
| 134 |
+
- name: terminology
|
| 135 |
+
dtype: float64
|
| 136 |
+
- name: length
|
| 137 |
+
dtype: float64
|
| 138 |
+
- name: deixis
|
| 139 |
+
dtype: float64
|
| 140 |
+
- name: faithfulness
|
| 141 |
+
dtype: float64
|
| 142 |
+
- name: non_redundancy
|
| 143 |
+
dtype: float64
|
| 144 |
+
- name: option_b-components
|
| 145 |
+
struct:
|
| 146 |
+
- name: style
|
| 147 |
+
dtype: float64
|
| 148 |
+
- name: terminology
|
| 149 |
+
dtype: float64
|
| 150 |
+
- name: length
|
| 151 |
+
dtype: float64
|
| 152 |
+
- name: deixis
|
| 153 |
+
dtype: float64
|
| 154 |
+
- name: faithfulness
|
| 155 |
+
dtype: float64
|
| 156 |
+
- name: non_redundancy
|
| 157 |
+
dtype: float64
|
| 158 |
- name: pair_id
|
| 159 |
dtype: string
|
| 160 |
- name: moment_id
|
|
|
|
| 167 |
dtype: string
|
| 168 |
- name: stratum
|
| 169 |
dtype: string
|
| 170 |
+
- name: cross_family
|
| 171 |
+
dtype: bool
|
| 172 |
+
- name: what_differs
|
| 173 |
+
dtype: string
|
| 174 |
- name: labels_pending
|
| 175 |
dtype: bool
|
| 176 |
+
splits:
|
| 177 |
+
- name: test_human
|
| 178 |
+
num_bytes: 1400519
|
| 179 |
+
num_examples: 400
|
| 180 |
+
download_size: 470863
|
| 181 |
+
dataset_size: 1400519
|
| 182 |
+
---
|
| 183 |
|
| 184 |
+
# Dataset Card for AD4Edu Preferences
|
| 185 |
|
| 186 |
Preference pairs over **audio descriptions (AD) of slide-based lecture videos**, for blind and
|
| 187 |
+
low-vision (BLV) students. Each pair is two candidate ADs for the same lecture *moment*; the task is to
|
| 188 |
+
say which better serves a BLV listener under the project's **45-rule** lecture-AD standard
|
| 189 |
+
(`rules_for_slides.yaml` v1.5; six categories: style, terminology, length, deixis, faithfulness,
|
| 190 |
+
non-redundancy).
|
| 191 |
|
| 192 |
**PRIVATE, and derived from copyrighted lecture video.** Do not redistribute. Keyframe images are
|
| 193 |
+
not included here; the moment context in `prompt` is text (transcript window, slide OCR, an on-screen
|
| 194 |
+
description, moment type, pause).
|
| 195 |
+
|
| 196 |
+
## Shape
|
| 197 |
+
|
| 198 |
+
The layout follows [`argilla/ultrafeedback-binarized-preferences-cleaned`](https://huggingface.co/datasets/argilla/ultrafeedback-binarized-preferences-cleaned),
|
| 199 |
+
adjusted to this corpus. `train_prefs` (config `default`):
|
| 200 |
|
| 201 |
+
| column | what it is |
|
| 202 |
+
|---|---|
|
| 203 |
+
| `source` | `natural_judged` (two families' candidates, ordered by a video judge in both presentation orders) or `constructed_counterfactual` (one axis of a compliant AD perturbed deliberately) |
|
| 204 |
+
| `prompt` | the moment as the SFT user turn renders it text-only: `Moment type`, `Lecturer said`, `On-screen text (OCR)`, `What is on screen`, `Pause available` |
|
| 205 |
+
| `prompt_id` | `sha256(prompt)`; equals the `text_only` config's `train_gen` `prompt_id` for the same moment on `Psychias/AD4Edu-SFT`, so the two datasets join |
|
| 206 |
+
| `chosen`, `rejected` | `[{user: prompt}, {assistant: AD text}]`; the assistant turn is the bare AD, which is what `train_dpo.py` trains on |
|
| 207 |
+
| `chosen-model`, `rejected-model` | the family that wrote each side (table below). On constructed pairs both name the base family: the rejected text is a perturbation of the chosen one, described by `axis` and `what_differs` |
|
| 208 |
+
| `chosen-rating`, `rejected-rating` | **ADEduEval/Core mechanical compliance** of each side, in [0, 1] -- see the next section before using them |
|
| 209 |
+
| `chosen-components`, `rejected-components` | the six category scores behind the rating; `null` = the tier abstains on that category |
|
| 210 |
+
| `pair_id`, `moment_id`, `lecture`, `type`, `axis`, `label_source`, `cross_family`, `what_differs` | provenance; `pair_id` is the foreign key the judge cache and the rater package hang off |
|
| 211 |
|
| 212 |
+
`test_human` (config `test_human`) has the same columns with `option_a` / `option_b` in place of
|
| 213 |
+
`chosen` / `rejected`, `stratum`, and `labels_pending`.
|
| 214 |
|
| 215 |
```python
|
| 216 |
from datasets import load_dataset
|
| 217 |
+
train = load_dataset("Psychias/AD4Edu-Preferences", "default")["train_prefs"]
|
| 218 |
+
held = load_dataset("Psychias/AD4Edu-Preferences", "test_human")["test_human"]
|
| 219 |
```
|
| 220 |
|
| 221 |
+
## The ratings are a diagnostic, not the label
|
| 222 |
|
| 223 |
+
`chosen-rating` is **not** how `chosen` was decided. The direction comes from `label_source`
|
| 224 |
+
(`judge` or `construction`); the rating is `ADEduEvalV2Fixed(use_models=False)` with style v2 --
|
| 225 |
+
tier **C-mech** of ADEduEval 1.0 (`docs/BENCHMARK.md` §1) -- computed on each side by
|
| 226 |
+
`ad4edu.validation.adedueval_eval.rate_pairs` and reproducible with
|
| 227 |
+
`python scripts/eval_adedueval.py hf --repo Psychias/AD4Edu-Preferences --check-shipped`.
|
| 228 |
|
| 229 |
+
> **ADEduEval/Core mechanical tier is a compliance DIAGNOSTIC (which rule does this AD break). FLAG-HARD on the shortcut gate: a slide-OCR readout beats a real AD on 99.4% of moments (analysis/degenerate_probe.md). It never ranks systems and never selects a checkpoint (docs/BENCHMARK.md sections 1 and 3). These components ARE the RL reward's rules term, so rl_guard refuses them on every reward-optimised arm.**
|
|
|
|
|
|
|
|
|
|
| 230 |
|
| 231 |
+
Concretely: the mechanical tier scores style, terminology, length and deixis (deixis abstains on
|
| 232 |
+
lectures whose recording renders no cursor, `analysis/cache/cursor_capability.json`); faithfulness
|
| 233 |
+
and non-redundancy are `null` because they need the model-backed tier. It penalises defects and
|
| 234 |
+
obliges nothing, so a slide readout can outscore a real description. Do not filter or reweight pairs
|
| 235 |
+
by rating margin and expect a quality signal. What the rating does tell you, per axis, is how often
|
| 236 |
+
the *compliant* side of a controlled pair also scores higher -- measured at build time on this split,
|
| 237 |
+
strata before any aggregate:
|
| 238 |
+
|
| 239 |
+
## by axis (strata first)
|
| 240 |
|
| 241 |
+
| stratum | pairs | moments | directed | ties | unscored | agree on decided | agree, ties = fail |
|
| 242 |
+
|---|---|---|---|---|---|---|---|
|
| 243 |
+
| deixis | 219 | 114 | 219 | 25 | 0 | 82.5% [76.5, 87.2] (k=160, n=194) | 73.1% [66.8, 78.5] (k=160, n=219) |
|
| 244 |
+
| faithfulness | 145 | 93 | 145 | 28 | 0 | 90.6% [83.9, 94.7] (k=106, n=117) | 73.1% [65.4, 79.7] (k=106, n=145) |
|
| 245 |
+
| length | 191 | 117 | 191 | 2 | 0 | 88.9% [83.6, 92.6] (k=168, n=189) | 88.0% [82.6, 91.8] (k=168, n=191) |
|
| 246 |
+
| natural | 2501 | 759 | 2501 | 58 | 0 | 52.4% [50.5, 54.4] (k=1281, n=2443) | 51.2% [49.3, 53.2] (k=1281, n=2501) |
|
| 247 |
+
| non_redundancy | 170 | 106 | 170 | 17 | 0 | 43.8% [36.2, 51.7] (k=67, n=153) | 39.4% [32.4, 46.9] (k=67, n=170) |
|
| 248 |
+
| style | 149 | 92 | 149 | 4 | 0 | 82.8% [75.8, 88.0] (k=120, n=145) | 80.5% [73.4, 86.1] (k=120, n=149) |
|
| 249 |
+
| terminology | 165 | 99 | 165 | 100 | 0 | 50.8% [38.9, 62.5] (k=33, n=65) | 20.0% [14.6, 26.8] (k=33, n=165) |
|
| 250 |
|
| 251 |
+
## by type (strata first)
|
| 252 |
|
| 253 |
+
| stratum | pairs | moments | directed | ties | unscored | agree on decided | agree, ties = fail |
|
| 254 |
+
|---|---|---|---|---|---|---|---|
|
| 255 |
+
| animation | 67 | 19 | 67 | 5 | 0 | 62.9% [50.5, 73.8] (k=39, n=62) | 58.2% [46.3, 69.3] (k=39, n=67) |
|
| 256 |
+
| chart | 17 | 4 | 17 | 0 | 0 | 47.1% [26.2, 69.0] (k=8, n=17) | 47.1% [26.2, 69.0] (k=8, n=17) |
|
| 257 |
+
| figure | 288 | 63 | 288 | 22 | 0 | 69.9% [64.2, 75.1] (k=186, n=266) | 64.6% [58.9, 69.9] (k=186, n=288) |
|
| 258 |
+
| ink | 32 | 6 | 32 | 3 | 0 | 41.4% [25.5, 59.3] (k=12, n=29) | 37.5% [22.9, 54.7] (k=12, n=32) |
|
| 259 |
+
| math | 36 | 15 | 36 | 0 | 0 | 44.4% [29.5, 60.4] (k=16, n=36) | 44.4% [29.5, 60.4] (k=16, n=36) |
|
| 260 |
+
| pointing | 2033 | 418 | 2033 | 165 | 0 | 59.8% [57.6, 62.0] (k=1117, n=1868) | 54.9% [52.8, 57.1] (k=1117, n=2033) |
|
| 261 |
+
| reveal | 183 | 62 | 183 | 3 | 0 | 40.6% [33.7, 47.9] (k=73, n=180) | 39.9% [33.1, 47.1] (k=73, n=183) |
|
| 262 |
+
| slide | 850 | 221 | 850 | 35 | 0 | 56.3% [52.9, 59.7] (k=459, n=815) | 54.0% [50.6, 57.3] (k=459, n=850) |
|
| 263 |
+
| table | 21 | 7 | 21 | 1 | 0 | 75.0% [53.1, 88.8] (k=15, n=20) | 71.4% [50.0, 86.2] (k=15, n=21) |
|
| 264 |
+
| transition | 13 | 2 | 13 | 0 | 0 | 76.9% [49.7, 91.8] (k=10, n=13) | 76.9% [49.7, 91.8] (k=10, n=13) |
|
| 265 |
+
|
| 266 |
+
**Aggregate, pooling axes ['deixis', 'faithfulness', 'length', 'natural', 'non_redundancy', 'style', 'terminology'] and types ['animation', 'chart', 'figure', 'ink', 'math', 'pointing', 'reveal', 'slide', 'table', 'transition']:**
|
| 267 |
+
|
| 268 |
+
| stratum | pairs | moments | directed | ties | unscored | agree on decided | agree, ties = fail |
|
| 269 |
+
|---|---|---|---|---|---|---|---|
|
| 270 |
+
| ALL | 3540 | 808 | 3540 | 234 | 0 | 58.5% [56.8, 60.2] (k=1935, n=3306) | 54.7% [53.0, 56.3] (k=1935, n=3540) |
|
| 271 |
+
|
| 272 |
+
|
| 273 |
+
## Label provenance -- read before using `train_prefs`
|
| 274 |
+
|
| 275 |
+
| `label_source` | n | how the direction is known |
|
| 276 |
|---|---|---|
|
| 277 |
| `construction` | 1039 | one axis perturbed deliberately; direction true by construction, no judge involved |
|
| 278 |
| `judge` | 2501 | Gemini judged a 16 s clip in **both** presentation orders and agreed with itself |
|
| 279 |
|
| 280 |
+
The judge was measured before use (figures as recorded at the v20 build, pooled over both judge
|
| 281 |
+
runs, n = 2,961 pairs; `analysis/j1_results.md`): **14.7 % of verdicts reversed under order swap**
|
| 282 |
+
[13.5, 16.0] and are excluded, which is why both orders are mandatory; **the longer side wins
|
| 283 |
+
56.9 %** [54.6, 59.2] of cross-family decided pairs (+6.9 pp), small but real, and it will bias any
|
| 284 |
+
length-sensitive comparison these labels feed; the judge's own family wins 54.9 % against Claude's
|
| 285 |
+
67.0 %, so self-preference is not borne out. Pairs the judge could not resolve (flips, ties,
|
| 286 |
+
unparsed) are **not** here. These caveats apply to the 2501 judge-labelled pairs; the 1039
|
| 287 |
+
construction-backed pairs carry neither. 53 natural pairs are same-family (`cross_family: false`).
|
| 288 |
|
| 289 |
+
Model columns carry the repo's **family key** (the unit `cross_family` is defined on and the suffix of every `output_id`); the provider id behind each key:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 290 |
|
| 291 |
+
| family key | provider | model id |
|
| 292 |
+
|---|---|---|
|
| 293 |
+
| `claude-sonnet-4.6` | openrouter | `anthropic/claude-sonnet-4.6` |
|
| 294 |
+
| `gemini-3.1-pro` | gemini | `gemini-3.1-pro-preview` |
|
| 295 |
+
| `gemini-3.1-pro-or` | openrouter | `google/gemini-3.1-pro-preview` |
|
| 296 |
+
| `gpt-5.5` | openrouter | `openai/gpt-5.5` |
|
| 297 |
+
| `qwen3-vl-235b` | openrouter | `qwen/qwen3-vl-235b-a22b-instruct` |
|
| 298 |
|
| 299 |
+
## Splits
|
| 300 |
+
|
| 301 |
+
| config / split | pairs | moments | labelled? | what it is |
|
| 302 |
+
|---|---|---|---|---|
|
| 303 |
+
| `default` / `train_prefs` | 3540 | 808 | yes (machine) | binarised `chosen` / `rejected`, for DPO; `train_eligible` rows of `dpo_train_v20.jsonl` |
|
| 304 |
+
| `test_human` / `test_human` | 400 | 271 | **no** | held out for blind/low-vision human annotation; strata {'natural_flip': 50, 'controlled': 250, 'natural_decided': 100} |
|
| 305 |
+
|
| 306 |
+
`test_human` ships **without a winner on purpose, and carries `labels_pending: true` to say so.**
|
| 307 |
+
The labels are not missing, they are *pending*: these are the pairs going to blind/low-vision raters,
|
| 308 |
+
and putting a machine label on them would substitute a guess for the measurement the split exists to
|
| 309 |
+
make. If you need labelled preference data, use `train_prefs`. The two options carry the same rating
|
| 310 |
+
and model columns as the training split, so the diagnostic is available on both sides without a label.
|
| 311 |
+
|
| 312 |
+
The two splits are **moment-disjoint**, not merely pair-disjoint: a moment's clip, transcript window
|
| 313 |
+
and references are shared across all its pairs, so effective n is moments. 0 eligible rows
|
| 314 |
+
had no direction and were dropped. The raw `dpo_train_v20.jsonl` (every row, with the churn flags that
|
| 315 |
+
make `train_eligible` inspectable) and `human_val_400.jsonl` ship beside the parquet.
|
| 316 |
|
| 317 |
+
Course-level holdout ("C1"): MATHEMATICS, DENTAL and MACHINE LEARNING are held out whole, derived
|
| 318 |
+
from the course map (`validation.split.is_c1_test`), with the reviewed lecture-level exceptions in
|
| 319 |
+
`TRAIN_OVERRIDE`.
|
| 320 |
|
| 321 |
+
## Changes from the v20 release
|
| 322 |
|
| 323 |
+
* `chosen` / `rejected` are message lists, not strings; `prompt_id` added.
|
| 324 |
+
* `chosen-model` / `rejected-model` (family keys) and `chosen-rating` / `rejected-rating` (+ components) added.
|
| 325 |
+
* `source` renamed from stage codes (`p3` / `p4`) to `natural_judged` / `constructed_counterfactual`.
|
| 326 |
+
* `test_human` is its own config (its columns differ from `train_prefs`); it gained option ratings and models.
|
| 327 |
+
* The standard is 45 rules (the v20 card said 42).
|
| 328 |
|
| 329 |
## Citation
|
| 330 |
|
data/test_human-00000-of-00001.parquet
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0975d4d14b2fa38f31a9d6c5e81537289be2ddba60198b7601e3c7da6976244e
|
| 3 |
+
size 470863
|
data/train_prefs-00000-of-00001.parquet
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d7f85684dbe41caa58cfc20038dc00fc7b908209013d22c586c6a9b68a807e17
|
| 3 |
+
size 1605312
|