broader-vids 0.1.11 → 0.1.12
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/docs/REFERENCE.md
CHANGED
|
@@ -1321,13 +1321,17 @@ Continuation preserves an explicit deadline and all cumulative request, media, a
|
|
|
1321
1321
|
|
|
1322
1322
|
YouTube metadata retrieval uses one active watch-page request and starts at least one second apart.
|
|
1323
1323
|
Bilibili metadata retrieval uses at most two concurrent requests and starts at least one second apart.
|
|
1324
|
-
|
|
1325
|
-
These
|
|
1326
|
-
|
|
1327
|
-
|
|
1324
|
+
Both platforms retry temporary metadata failures: HTTP 408, 500, 502, 503, or 504, timeouts, connection resets, and incomplete responses.
|
|
1325
|
+
These failures and Bilibili API code `-404` share one initial attempt and up to three retries per candidate.
|
|
1326
|
+
YouTube's identity-repair retry uses the same allowance. It does not start a separate retry cycle.
|
|
1327
|
+
Retries use the same session, pacing, and request allowance.
|
|
1328
|
+
Recorded failures remain counted when a paused round continues. A new round has its own count.
|
|
1329
|
+
A successful retry returns normal metadata without a score deduction.
|
|
1330
|
+
After retry exhaustion, collection skips and reports that candidate, then fills its lane from other candidates.
|
|
1328
1331
|
Saved eligible candidates and unfinished batches remain available on continuation.
|
|
1329
1332
|
This result does not prove permanent deletion and does not assign a quality score or preference rating.
|
|
1330
|
-
Login errors, access challenges, malformed metadata, and mismatched video identities still pause collection.
|
|
1333
|
+
Login errors, rate limits, access challenges, malformed metadata, and mismatched video identities still pause collection.
|
|
1334
|
+
The engine does not retry HTTP 401, 403, 412, or 429 automatically.
|
|
1331
1335
|
Workers retain server-issued session cookies in memory.
|
|
1332
1336
|
Pacing waits count toward an optional saved time allowance.
|
|
1333
1337
|
The [yt-dlp request-rate guidance](https://github.com/yt-dlp/yt-dlp/wiki/Extractors#common-youtube-errors) recommends longer waits between downloads when rate limits occur.
|
package/docs/SCORING.md
CHANGED
|
@@ -52,22 +52,22 @@ Bilibili additionally requires coins/views >= 0.001 and favorites/views >= 0.002
|
|
|
52
52
|
Missing required metrics fail admission.
|
|
53
53
|
These hard floors use raw rates before the smoothing described below.
|
|
54
54
|
|
|
55
|
-
For metric count
|
|
55
|
+
For metric count $c$, views $v$, and platform baseline $b$, the adjusted rate is
|
|
56
56
|
|
|
57
|
-
|
|
57
|
+
$$
|
|
58
58
|
\widehat r=\frac{c+1000b}{v+1000}.
|
|
59
|
-
|
|
59
|
+
$$
|
|
60
60
|
|
|
61
61
|
The baseline is the middle sorted rate when at least ten valid platform observations exist.
|
|
62
62
|
Otherwise, initial baselines are 0.03 for likes, 0.01 for coins, and 0.02 for favorites.
|
|
63
|
-
The implementation limits the baseline below by
|
|
63
|
+
The implementation limits the baseline below by $10^{-6}$.
|
|
64
64
|
|
|
65
65
|
The metric score is
|
|
66
66
|
|
|
67
|
-
|
|
67
|
+
$$
|
|
68
68
|
e=100\sigma\left(\operatorname{clip}\left(\log\frac{\max(\widehat r,10^{-12})}{b},-8,8\right)\right),
|
|
69
69
|
\qquad \sigma(z)=\frac{1}{1+e^{-z}}.
|
|
70
|
-
|
|
70
|
+
$$
|
|
71
71
|
|
|
72
72
|
The rate-score helper assigns 50 to a missing count or zero denominator.
|
|
73
73
|
Such entries fail the hard admission floors and cannot enter a new recommendation shortlist.
|
|
@@ -147,16 +147,16 @@ Creator reputation does not guarantee a high grade.
|
|
|
147
147
|
Existing scorecards remain stored. New work and cache reuse require rubric 5.
|
|
148
148
|
|
|
149
149
|
An assessed criterion has an integer grade from zero to four.
|
|
150
|
-
Its normalized value is
|
|
150
|
+
Its normalized value is $u_j=\mathrm{grade}_j/4$.
|
|
151
151
|
An unknown criterion uses a prior mean of 0.5 and a separate missingness predictor.
|
|
152
152
|
Its stored grade remains unknown.
|
|
153
153
|
It does not inherit a high grade from another criterion.
|
|
154
154
|
|
|
155
|
-
For duration
|
|
155
|
+
For duration $t$ seconds, the derived initial quality index is
|
|
156
156
|
|
|
157
|
-
|
|
157
|
+
$$
|
|
158
158
|
d=\min(t/3600,1),\qquad Q_0=95\sum_j w_j u_j+5d.
|
|
159
|
-
|
|
159
|
+
$$
|
|
160
160
|
|
|
161
161
|
Length contributes at most five points.
|
|
162
162
|
The contribution rises to one hour and then remains constant.
|
|
@@ -168,7 +168,7 @@ It ignores retrieval origin and accepts only medium- or high-confidence matches.
|
|
|
168
168
|
Accepted matches supply canonical broad-root features for learned lookup.
|
|
169
169
|
Assessed topic labels remain unchanged.
|
|
170
170
|
|
|
171
|
-
The initial interest probability
|
|
171
|
+
The initial interest probability $I_0$ is 0.75 for a domain related to the user's stated interests.
|
|
172
172
|
It is 0.35 for a known unrelated domain and 0.50 when domains are unknown.
|
|
173
173
|
Search-domain labels describe retrieval context.
|
|
174
174
|
They do not prove the video's semantic topic.
|
|
@@ -191,10 +191,10 @@ Current ratings use categorical topic, subject, depth, and quality responses.
|
|
|
191
191
|
The quality model factors its categorical likelihood into logistic events.
|
|
192
192
|
Each event uses a predictor vector and a proper Gaussian prior:
|
|
193
193
|
|
|
194
|
-
|
|
194
|
+
$$
|
|
195
195
|
y_i\sim\operatorname{Bernoulli}(\sigma(o_i+x_i^T\beta)),
|
|
196
196
|
\qquad \beta_j\sim N(0,s_j^2).
|
|
197
|
-
|
|
197
|
+
$$
|
|
198
198
|
|
|
199
199
|
Numerical clipping keeps offsets finite.
|
|
200
200
|
It does not limit the final learned probability.
|
|
@@ -216,20 +216,20 @@ Model version 8 retains its portable parameter groups for exact preference trans
|
|
|
216
216
|
|
|
217
217
|
The initial weighted score is
|
|
218
218
|
|
|
219
|
-
|
|
219
|
+
$$
|
|
220
220
|
B=0.60Q+0.25I+0.05E+0.05R+0.05D.
|
|
221
|
-
|
|
221
|
+
$$
|
|
222
222
|
|
|
223
223
|
All components use a scale from zero to 100.
|
|
224
224
|
Without structured feedback, quality and interest use their initial priors.
|
|
225
225
|
The next section defines the structured replacements.
|
|
226
|
-
|
|
227
|
-
For age
|
|
228
|
-
A missing publication date gives
|
|
226
|
+
$E$ is community response.
|
|
227
|
+
For age $a$ in years, recency is $R=100/(1+a/10)$.
|
|
228
|
+
A missing publication date gives $R=50$.
|
|
229
229
|
|
|
230
|
-
Creator diversity is
|
|
231
|
-
Here,
|
|
232
|
-
The value
|
|
230
|
+
Creator diversity is $D=100/(1+n+0.1h)$.
|
|
231
|
+
Here, $n$ counts this creator in the developing list.
|
|
232
|
+
The value $h$ counts the creator in recent recommendation exposures.
|
|
233
233
|
Randomness enters discovery rather than adding a random final-score bonus.
|
|
234
234
|
This score component measures creator diversity.
|
|
235
235
|
Screening policy 5 also applies a separate diversity rule when it constructs each platform list.
|
|
@@ -246,7 +246,7 @@ It adds no model calls. It cannot guarantee semantic uniqueness when labels or q
|
|
|
246
246
|
Different videos about the same subject can therefore recur.
|
|
247
247
|
A subject match alone does not establish repeated content or equal explanatory depth.
|
|
248
248
|
|
|
249
|
-
Current profiles rank with
|
|
249
|
+
Current profiles rank with $B$, using learned structured quality and interest.
|
|
250
250
|
The initial component weights remain a ranking policy.
|
|
251
251
|
An empty internal overall model does not change this score.
|
|
252
252
|
Historical snapshots can retain an older overall adjustment.
|
|
@@ -278,15 +278,15 @@ The user can remove an exclusion.
|
|
|
278
278
|
### Topic and broad-subject interest
|
|
279
279
|
|
|
280
280
|
Each user-selected topic or subject uses a separate Dirichlet categorical model.
|
|
281
|
-
For initial mean
|
|
281
|
+
For initial mean $b$, its Low/Medium/High prior is
|
|
282
282
|
|
|
283
|
-
|
|
283
|
+
$$
|
|
284
284
|
\alpha=2\left((1-b)^2,\ 2b(1-b),\ b^2\right).
|
|
285
|
-
|
|
285
|
+
$$
|
|
286
286
|
|
|
287
|
-
Use
|
|
288
|
-
Current distinct-video category counts
|
|
289
|
-
The interest index is
|
|
287
|
+
Use $b=0.75$ for an onboarding-related topic or subject and $b=0.35$ otherwise.
|
|
288
|
+
Current distinct-video category counts $n$ give posterior $\operatorname{Dirichlet}(\alpha+n)$.
|
|
289
|
+
The interest index is $100[P(\mathrm{High})+0.5P(\mathrm{Medium})]$.
|
|
290
290
|
|
|
291
291
|
The earliest explicit topic or depth action anchors the topic's parent domains.
|
|
292
292
|
Edits and clears preserve that anchor.
|
|
@@ -312,25 +312,26 @@ Custom topic labels transfer only to matching normalized labels.
|
|
|
312
312
|
|
|
313
313
|
Depth uses an exact categorical model for each observed topic and audience level.
|
|
314
314
|
The responses are want deeper, suitable, and want lighter.
|
|
315
|
-
Their Dirichlet prior is
|
|
316
|
-
For an observed level,
|
|
315
|
+
Their Dirichlet prior is $(0.5,1,0.5)$.
|
|
316
|
+
For an observed level, $s=P(\mathrm{suitable})$ multiplies interest odds by $s/(1-s)$.
|
|
317
317
|
Its prior mean of 0.5 is neutral.
|
|
318
318
|
|
|
319
319
|
For a level without direct feedback, a separate directional model transfers information within that topic.
|
|
320
|
-
Encode introductory, intermediate, and advanced as
|
|
321
|
-
A latent preferred level has prior
|
|
322
|
-
With
|
|
320
|
+
Encode introductory, intermediate, and advanced as $l=0,1,2$.
|
|
321
|
+
A latent preferred level has prior $\mu\sim N(1,1.5^2)$.
|
|
322
|
+
With $w=0.75$, its likelihood is
|
|
323
323
|
|
|
324
|
-
|
|
324
|
+
$$
|
|
325
325
|
P(\mathrm{deeper}\mid\mu,l)=\sigma(\mu-l-w),
|
|
326
326
|
\quad P(\mathrm{lighter}\mid\mu,l)=\sigma(l-\mu-w),
|
|
327
|
-
|
|
328
|
-
|
|
327
|
+
$$
|
|
328
|
+
|
|
329
|
+
$$
|
|
329
330
|
P(\mathrm{suitable}\mid\mu,l)=1-P(\mathrm{deeper})-P(\mathrm{lighter}).
|
|
330
|
-
|
|
331
|
+
$$
|
|
331
332
|
|
|
332
333
|
Adaptive one-dimensional quadrature approximates this posterior.
|
|
333
|
-
The transfer adds
|
|
334
|
+
The transfer adds $\operatorname{logit}(s_{post})-\operatorname{logit}(s_{prior})$ to interest log odds.
|
|
334
335
|
Thus, no feedback gives no adjustment.
|
|
335
336
|
Want deeper shifts preference upward, while want lighter shifts it downward.
|
|
336
337
|
Direct categorical feedback takes precedence when that candidate level receives observations.
|
|
@@ -341,7 +342,7 @@ Its fixed noise width cannot represent every possible response distribution.
|
|
|
341
342
|
The exact categorical consistency result applies to directly observed topic-level groups.
|
|
342
343
|
No unconditional convergence claim applies to transferred scores at an unobserved level.
|
|
343
344
|
|
|
344
|
-
The engine clips depth suitability to
|
|
345
|
+
The engine clips depth suitability to $10^{-12}$ through $1-10^{-12}$ before conversion to log odds.
|
|
345
346
|
Depth does not change quality or domain interest.
|
|
346
347
|
Unknown audience levels retain feedback but do not adjust future scores.
|
|
347
348
|
The model supplies a cited level when the evidence supports one.
|
|
@@ -356,16 +357,16 @@ Duration does not establish audience level.
|
|
|
356
357
|
|
|
357
358
|
The quality model uses two logistic stages:
|
|
358
359
|
|
|
359
|
-
|
|
360
|
+
$$
|
|
360
361
|
a=P(\mathrm{Medium\ or\ High}),\quad b=P(\mathrm{High}\mid\mathrm{not\ Low}).
|
|
361
|
-
|
|
362
|
+
$$
|
|
362
363
|
|
|
363
|
-
The category probabilities are
|
|
364
|
+
The category probabilities are $(1-a,\ a(1-b),\ ab)$.
|
|
364
365
|
A Low response supplies only the first-stage event.
|
|
365
366
|
Medium and High responses also supply the conditional second-stage event.
|
|
366
367
|
This is one categorical likelihood, not duplicate votes or fractional Bernoulli labels.
|
|
367
368
|
|
|
368
|
-
The quality index is
|
|
369
|
+
The quality index is $Q=100[0.5P(\mathrm{Medium})+P(\mathrm{High})]$.
|
|
369
370
|
It represents expected ordinal utility, not a binary approval percentage.
|
|
370
371
|
Initial stage offsets use the quality prior when the retired binary model has no observations.
|
|
371
372
|
Learned corrections can cancel these offsets.
|
|
@@ -451,12 +452,12 @@ They continue to label local videos without adding their old ratings to training
|
|
|
451
452
|
|
|
452
453
|
For a logistic parameter vector, import uses the saved Gaussian approximation as a proper prior:
|
|
453
454
|
|
|
454
|
-
|
|
455
|
+
$$
|
|
455
456
|
\beta\sim N(m_0,\Sigma_0),\qquad
|
|
456
457
|
\log p(\beta\mid D_{\mathrm{new}})
|
|
457
458
|
=-\tfrac12(\beta-m_0)^\top\Sigma_0^{-1}(\beta-m_0)
|
|
458
459
|
+\sum_{i\in D_{\mathrm{new}}}\log p(y_i\mid x_i,\beta)+C.
|
|
459
|
-
|
|
460
|
+
$$
|
|
460
461
|
|
|
461
462
|
The implementation preserves off-diagonal correlations.
|
|
462
463
|
New coefficients receive the usual independent proper priors.
|
|
@@ -511,7 +512,7 @@ Larger models use matrix-free preconditioned conjugate gradients.
|
|
|
511
512
|
Prediction uses finite Gaussian quadrature.
|
|
512
513
|
These are numerical approximations, with conditional asymptotic justification rather than exact finite-sample posterior calculations.
|
|
513
514
|
|
|
514
|
-
For a fixed categorical group, the exact posterior is
|
|
515
|
+
For a fixed categorical group, the exact posterior is $\operatorname{Dirichlet}(\alpha+n)$.
|
|
515
516
|
If that group's distinct-video sample grows under stationary independent responses, empirical proportions converge and posterior variance tends to zero.
|
|
516
517
|
Each topic or level group needs its own coverage.
|
|
517
518
|
The ordinal model also requires positive second-stage frequency and well-specified conditional models.
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "broader-vids",
|
|
3
|
-
"version": "0.1.
|
|
3
|
+
"version": "0.1.12",
|
|
4
4
|
"description": "Codex and experimental Claude Code skills for local YouTube/Bilibili recommendations",
|
|
5
5
|
"license": "MIT",
|
|
6
6
|
"engines": {
|
|
@@ -11,7 +11,7 @@
|
|
|
11
11
|
},
|
|
12
12
|
"files": [
|
|
13
13
|
"bin/",
|
|
14
|
-
"python/broader_vids-0.1.
|
|
14
|
+
"python/broader_vids-0.1.12-*.whl",
|
|
15
15
|
"python/core-requirements.txt",
|
|
16
16
|
"python/evidence-requirements.txt",
|
|
17
17
|
"python/asr-requirements.txt",
|
|
@@ -48,6 +48,6 @@
|
|
|
48
48
|
"tag": "latest"
|
|
49
49
|
},
|
|
50
50
|
"broaderVidsRelease": {
|
|
51
|
-
"sourceCommit": "
|
|
51
|
+
"sourceCommit": "450455e1888fec57843762d8245cf69b335d43c5"
|
|
52
52
|
}
|
|
53
53
|
}
|
package/plugin.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
{
|
|
2
2
|
"$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
|
|
3
3
|
"name": "broader-vids",
|
|
4
|
-
"version": "0.1.
|
|
4
|
+
"version": "0.1.12",
|
|
5
5
|
"description": "Curated YouTube and Bilibili recommendations with bounded quality evidence and local watched/feedback state.",
|
|
6
6
|
"license": "MIT",
|
|
7
7
|
"extensions": {
|
package/python/{broader_vids-0.1.11-py3-none-any.whl → broader_vids-0.1.12-py3-none-any.whl}
RENAMED
|
index 126dca9..f2bdd99 100644
|
|
|
Binary file
|