Spaces:
Paused
Paused
Malina Claude Opus 4.8 commited on
Commit Β·
d0ff022
1
Parent(s): 51e7ae9
Fix diarization hub kwarg, yt-dlp cookies, stronger output formatting
Browse files- diarize: shim huggingface_hub.hf_hub_download/snapshot_download to translate the
removed use_auth_token kwarg β token, so pyannote 3.3.2 loads on hub 1.x
(fixes "Diarization unavailable: ... unexpected keyword argument 'use_auth_token'").
- fetch: support YouTube/Facebook cookies (YT_DLP_COOKIES[_B64] secret) + a real
browser UA, the actual fix for "Sign in to confirm you're not a bot / age" on
cloud IPs; still best-effort.
- synthesis/chat: enforce a strong "CRITICAL OUTPUT FORMAT" (headings, bullets,
numbered steps, tables for comparisons; no walls of text) for scannable answers.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- chat.py +5 -4
- diarize.py +21 -0
- fetch.py +40 -0
- synthesis.py +7 -7
chat.py
CHANGED
|
@@ -132,10 +132,11 @@ def _answer_from_context(messages: list, transcript: str):
|
|
| 132 |
f"Ground everything in the conversation and the stored transcript below; do not invent "
|
| 133 |
f"facts. Stay on the video: if asked something unrelated to it, briefly steer back. "
|
| 134 |
f"If the answer needs visual detail you don't have, say so and suggest re-analyzing. "
|
| 135 |
-
f"
|
| 136 |
-
f"
|
| 137 |
-
f"
|
| 138 |
-
f"text,
|
|
|
|
| 139 |
)
|
| 140 |
user = f"Stored video transcript:\n{transcript or '(none yet)'}\n\nConversation:\n{_render_history(messages)}"
|
| 141 |
if last.get("quote"):
|
|
|
|
| 132 |
f"Ground everything in the conversation and the stored transcript below; do not invent "
|
| 133 |
f"facts. Stay on the video: if asked something unrelated to it, briefly steer back. "
|
| 134 |
f"If the answer needs visual detail you don't have, say so and suggest re-analyzing. "
|
| 135 |
+
f"CRITICAL OUTPUT FORMAT β your response must be visually structured; assume the user is "
|
| 136 |
+
f"skimming, not reading word-for-word, so optimize for scannability. Use (whichever fit) "
|
| 137 |
+
f"Markdown headings, bullet lists, numbered steps, and tables when comparing items. "
|
| 138 |
+
f"Forbidden: large blocks of uninterrupted text, or any answer longer than ~5 lines "
|
| 139 |
+
f"without formatting. A trivial one-line answer may stay plain β don't pad it."
|
| 140 |
)
|
| 141 |
user = f"Stored video transcript:\n{transcript or '(none yet)'}\n\nConversation:\n{_render_history(messages)}"
|
| 142 |
if last.get("quote"):
|
diarize.py
CHANGED
|
@@ -59,6 +59,26 @@ def _shim_torchaudio(torchaudio) -> None:
|
|
| 59 |
_shimmed = True
|
| 60 |
|
| 61 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 62 |
def _get_pipeline():
|
| 63 |
global _pipeline, _failed
|
| 64 |
if _failed:
|
|
@@ -68,6 +88,7 @@ def _get_pipeline():
|
|
| 68 |
import torch
|
| 69 |
import torchaudio
|
| 70 |
_shim_torchaudio(torchaudio) # paper over torchaudio 2.11's removed I/O API
|
|
|
|
| 71 |
from pyannote.audio import Pipeline
|
| 72 |
# huggingface-hub renamed use_auth_token β token; support both.
|
| 73 |
try:
|
|
|
|
| 59 |
_shimmed = True
|
| 60 |
|
| 61 |
|
| 62 |
+
def _shim_hf_hub() -> None:
|
| 63 |
+
"""pyannote 3.3.2 internally calls hf_hub_download(use_auth_token=...) β a kwarg
|
| 64 |
+
that huggingface-hub 1.x removed (it's now `token`). Translate it so the gated
|
| 65 |
+
model still loads. Must run BEFORE `from pyannote.audio import Pipeline` so
|
| 66 |
+
pyannote binds the patched function."""
|
| 67 |
+
import huggingface_hub as _hf
|
| 68 |
+
for _name in ("hf_hub_download", "snapshot_download"):
|
| 69 |
+
_orig = getattr(_hf, _name, None)
|
| 70 |
+
if _orig is None or getattr(_orig, "_nina_shim", False):
|
| 71 |
+
continue
|
| 72 |
+
def _make(orig):
|
| 73 |
+
def _wrapped(*a, **k):
|
| 74 |
+
if "use_auth_token" in k:
|
| 75 |
+
k.setdefault("token", k.pop("use_auth_token"))
|
| 76 |
+
return orig(*a, **k)
|
| 77 |
+
_wrapped._nina_shim = True
|
| 78 |
+
return _wrapped
|
| 79 |
+
setattr(_hf, _name, _make(_orig))
|
| 80 |
+
|
| 81 |
+
|
| 82 |
def _get_pipeline():
|
| 83 |
global _pipeline, _failed
|
| 84 |
if _failed:
|
|
|
|
| 88 |
import torch
|
| 89 |
import torchaudio
|
| 90 |
_shim_torchaudio(torchaudio) # paper over torchaudio 2.11's removed I/O API
|
| 91 |
+
_shim_hf_hub() # paper over huggingface-hub's use_auth_token removal
|
| 92 |
from pyannote.audio import Pipeline
|
| 93 |
# huggingface-hub renamed use_auth_token β token; support both.
|
| 94 |
try:
|
fetch.py
CHANGED
|
@@ -7,6 +7,7 @@
|
|
| 7 |
# website always offers file upload as the fallback. We pull a low-res stream (NINA
|
| 8 |
# only needs audio + sampled frames) and cap the duration.
|
| 9 |
|
|
|
|
| 10 |
import os
|
| 11 |
import tempfile
|
| 12 |
|
|
@@ -15,6 +16,41 @@ from auth import verify_usage_token
|
|
| 15 |
MAX_FETCH_SECONDS = 45 * 60 # reject videos longer than 45 min
|
| 16 |
# Prefer a small β€480p stream (+ best audio); NINA downscales/ samples anyway.
|
| 17 |
YT_DLP_FORMAT = "bv*[height<=480]+ba/b[height<=480]/best"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
|
| 20 |
def fetch_video(url: str, token: str = ""):
|
|
@@ -43,7 +79,11 @@ def fetch_video(url: str, token: str = ""):
|
|
| 43 |
"socket_timeout": 30,
|
| 44 |
"retries": 2,
|
| 45 |
"concurrent_fragment_downloads": 4,
|
|
|
|
| 46 |
}
|
|
|
|
|
|
|
|
|
|
| 47 |
|
| 48 |
try:
|
| 49 |
with yt_dlp.YoutubeDL(opts) as ydl:
|
|
|
|
| 7 |
# website always offers file upload as the fallback. We pull a low-res stream (NINA
|
| 8 |
# only needs audio + sampled frames) and cap the duration.
|
| 9 |
|
| 10 |
+
import base64
|
| 11 |
import os
|
| 12 |
import tempfile
|
| 13 |
|
|
|
|
| 16 |
MAX_FETCH_SECONDS = 45 * 60 # reject videos longer than 45 min
|
| 17 |
# Prefer a small β€480p stream (+ best audio); NINA downscales/ samples anyway.
|
| 18 |
YT_DLP_FORMAT = "bv*[height<=480]+ba/b[height<=480]/best"
|
| 19 |
+
# A real browser UA helps a little against bot checks; cookies are the actual fix.
|
| 20 |
+
_UA = ("Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
|
| 21 |
+
"(KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36")
|
| 22 |
+
|
| 23 |
+
_cookie_path = None
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
def _cookiefile():
|
| 27 |
+
"""YouTube/Facebook block datacenter IPs ('confirm you're not a bot' / age gate).
|
| 28 |
+
Set a Space secret with an exported Netscape cookies.txt to authenticate yt-dlp:
|
| 29 |
+
β’ YT_DLP_COOKIES_B64 β base64 of the cookies.txt (preferred; survives newlines), or
|
| 30 |
+
β’ YT_DLP_COOKIES β the raw cookies.txt contents.
|
| 31 |
+
Returns a path to a temp cookie file, or None when no cookies are configured."""
|
| 32 |
+
global _cookie_path
|
| 33 |
+
if _cookie_path and os.path.exists(_cookie_path):
|
| 34 |
+
return _cookie_path
|
| 35 |
+
raw = None
|
| 36 |
+
b64 = os.getenv("YT_DLP_COOKIES_B64")
|
| 37 |
+
if b64:
|
| 38 |
+
try:
|
| 39 |
+
raw = base64.b64decode(b64).decode("utf-8", "replace")
|
| 40 |
+
except Exception:
|
| 41 |
+
raw = None
|
| 42 |
+
if raw is None:
|
| 43 |
+
raw = os.getenv("YT_DLP_COOKIES")
|
| 44 |
+
if not raw or not raw.strip():
|
| 45 |
+
return None
|
| 46 |
+
try:
|
| 47 |
+
p = os.path.join(tempfile.gettempdir(), "nina_yt_cookies.txt")
|
| 48 |
+
with open(p, "w", encoding="utf-8") as f:
|
| 49 |
+
f.write(raw)
|
| 50 |
+
_cookie_path = p
|
| 51 |
+
return p
|
| 52 |
+
except Exception:
|
| 53 |
+
return None
|
| 54 |
|
| 55 |
|
| 56 |
def fetch_video(url: str, token: str = ""):
|
|
|
|
| 79 |
"socket_timeout": 30,
|
| 80 |
"retries": 2,
|
| 81 |
"concurrent_fragment_downloads": 4,
|
| 82 |
+
"http_headers": {"User-Agent": _UA},
|
| 83 |
}
|
| 84 |
+
cookies = _cookiefile()
|
| 85 |
+
if cookies:
|
| 86 |
+
opts["cookiefile"] = cookies
|
| 87 |
|
| 88 |
try:
|
| 89 |
with yt_dlp.YoutubeDL(opts) as ydl:
|
synthesis.py
CHANGED
|
@@ -124,13 +124,13 @@ def _detect_language(text: str) -> str:
|
|
| 124 |
# which the chat renders) instead of always emitting one prose block. Shared by the
|
| 125 |
# text and vision synthesis prompts.
|
| 126 |
_FORMAT_GUIDE = (
|
| 127 |
-
"FORMAT β
|
| 128 |
-
"
|
| 129 |
-
"
|
| 130 |
-
"
|
| 131 |
-
"
|
| 132 |
-
"
|
| 133 |
-
"
|
| 134 |
"Do NOT begin lines with timestamps. End with exactly one line: Sources: [m:ss], [m:ss]."
|
| 135 |
)
|
| 136 |
|
|
|
|
| 124 |
# which the chat renders) instead of always emitting one prose block. Shared by the
|
| 125 |
# text and vision synthesis prompts.
|
| 126 |
_FORMAT_GUIDE = (
|
| 127 |
+
"CRITICAL OUTPUT FORMAT β your response must be visually structured. Assume the user "
|
| 128 |
+
"is skimming, not reading word-for-word; optimize for scannability.\n"
|
| 129 |
+
"Required (use whichever fit the answer): Markdown headings, bullet lists, numbered "
|
| 130 |
+
"steps, and tables (Markdown `| col | col |`) when comparing items.\n"
|
| 131 |
+
"Forbidden: large blocks of uninterrupted text; any answer longer than ~5 lines without "
|
| 132 |
+
"formatting.\n"
|
| 133 |
+
"(A trivial one-line factual answer may stay plain β don't pad it into a table.)\n"
|
| 134 |
"Do NOT begin lines with timestamps. End with exactly one line: Sources: [m:ss], [m:ss]."
|
| 135 |
)
|
| 136 |
|