yumoqing/pipeline-platform
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(platform): 能力守卫歧义处理——'视频'字在t2v描述必现不算输入声明;音频单输入不自动纠正只警告

Browse Source
This commit is contained in:
yumoqing 2026-09-06 14:58:38 +08:00
parent b2ecf54f43
commit 6b306566b2
1 changed files with 24 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

43
pipeline_platform/platform_ability.py
View File

@ -486,6 +486,8 @@ def _capability_guard(capability, description):
保守策略：只在「纯文本能力（t2v/t2i/t2t）+ 描述含媒体输入/参考声明」这种
铁证矛盾时纠正；其他不一致只警告不改（避免误伤描述措辞不规范的正确配置）。
歧义处理："视频"二字在 t2v/i2v 描述里几乎必现（输出侧），不算输入声明；
输入声明判据 = 参考词（参考生/全能参考/四模态等）或图像/音频词或输入媒体≥2类。
返回 (corrected_cap 或 None, warning 或 None)。
"""
desc = description or ''
@ -493,26 +495,29 @@ def _capability_guard(capability, description):
return None, None
has_ref = any(w in desc for w in _DESC_REF_WORDS)
kinds = [k for k, words in _DESC_MEDIA_WORDS.items() if any(w in desc for w in words)]
# 「输出 视频」的"视频"不算输入声明——t2v 描述几乎都带"视频"二字。
# 输入声明判据：参考词，或图像/音频词（这两个不会出现在纯文本生成的描述里），
# 或"视频"以外多模态组合。
input_kinds = [k for k in kinds if k != 'video']
if has_ref or len(input_kinds) >= 1 or (has_ref is False and len(kinds) >= 2 and 'video' in kinds and input_kinds):
out_word = '视频' if capability == 't2v' else ('图像' if capability == 't2i' else '')
if capability == 't2t':
return None, ("描述含媒体输入声明（%s）但能力是 t2t——请人工核对能力分类"
% '/'.join(kinds))
expected = 'r2v' if (has_ref or len(kinds) >= 2) else (
'i2v' if 'image' in input_kinds and out_word == '视频' else
('i2i' if 'image' in input_kinds and out_word == '图像' else None))
if expected and expected != capability:
return expected, ("能力自动纠正 %s→%s：描述声明媒体输入（%s%s），"
"纯文本能力与之矛盾（提取照示例误判，守卫按正文模态纠正）"
% (capability, expected, '/'.join(kinds),
'，含参考声明' if has_ref else ''))
return None, ("描述含媒体输入声明（%s）但能力是 %s——无法确定目标能力，"
"请人工核对（可用 overrides.model_capability 纠正）"
% ('/'.join(kinds), capability))
if not (has_ref or input_kinds):
return None, None
if capability == 't2t':
return None, ("描述含媒体输入声明（%s%s）但能力是 t2t——请人工核对能力分类"
% ('/'.join(kinds), '，含参考声明' if has_ref else ''))
out_word = '视频' if capability == 't2v' else '图像'
expected = None
if has_ref:
expected = 'r2v' if capability == 't2v' else 'i2i'
elif len(input_kinds) >= 2:
expected = 'r2v' if capability == 't2v' else 'i2i'
elif 'image' in input_kinds:
expected = 'i2v' if capability == 't2v' else 'i2i'
elif 'audio' in input_kinds:
# 音频输入语义不定（r2v 音频参考 / asr / tts 反向），不敢自动纠正
return None, ("描述含音频输入声明但能力是 %s——无法确定目标能力，"
"请人工核对（可用 overrides.model_capability 纠正）" % capability)
if expected and expected != capability:
return expected, ("能力自动纠正 %s→%s：描述声明媒体输入（%s%s），"
"纯文本能力与之矛盾（提取照示例误判，守卫按正文模态纠正）"
% (capability, expected, '/'.join(kinds),
'，含参考声明' if has_ref else ''))
return None, None