CTGT Finds Distilling DeepSeek into GPT-OSS Does Not Transfer Censorship, Releases LineageEval Framework
English summary
CTGT used DeepSeek V4 Flash as a teacher to distill GPT-OSS-120B for finance tasks, achieving 83.61% on FinanceReasoning at an 8k token budget, outperforming Kimi K3 and Inkling. They measured censorship transfer with 152 matched political prompt pairs scored by four LLM judges; the teacher showed a +45.45 point gap (7 SD from chance) in avoiding China‑sensitive topics, but all distilled students stayed within 1 point of their American base model. The censorship from the Chinese teacher did not transfer. CTGT released the open‑evaluation framework LineageEval, open‑weight 20B finance model, and a playground for side‑by‑side testing. They plan to extend the study using Chinese‑lineage base models like Qwen.
Chinese summary
CTGT 采用 DeepSeek V4 Flash 作为教师，将 GPT-OSS-120B 蒸馏用于金融任务，在 8k 令牌预算下取得 FinanceReasoning 上 83.61% 的分数，超越 Kimi K3 和 Inkling。他们通过 152 对匹配的政治提示，由四位 LLM 评委评分，测量审查传递性；教师模型在中国敏感话题上的回避偏差达 +45.45 分（距偶然 7 个标准差），而所有蒸馏学生与其美国基座的行为差距均小于 1 分，审查未传递。CTGT 开源了评估框架 LineageEval、20B 金融模型权重以及一个并行测试平台。下一步计划在 Qwen 等中国血统的基础上开展类似实验。
Key points
Distilled GPT-OSS-120B achieved 83.61% on FinanceReasoning at 8k tokens, beating Kimi K3 (81.93%) and Inkling (65.13%).
蒸馏后的 GPT-OSS-120B 在 8k 令牌的 FinanceReasoning 上取得 83.61%，优于 Kimi K3（81.93%）和 Inkling（65.13%）。
Censorship did not transfer: the DeepSeek teacher showed a +45.45 point gap (~7 SD) on politically sensitive pairs, while all distilled variants remained within 1 point of their American base model.
审查未传递：DeepSeek 教师模型在政治敏感问题上表现出 +45.45 分的偏差（约 7 个标准差），而所有蒸馏版本与其美国基座的行为相差均在 1 分以内。
LineageEval, an open evaluation framework with 152 matched prompts, LLM judging protocol, and human validation (r=0.948), was released to enable transparent censorship transfer measurement.
开源了 LineageEval 评估框架，包含 152 对匹配提示、LLM 评判协议及人工验证（相关系数 0.948），以支持透明的审查传递性测量。
Open-weight 20B finance model and a no‑auth playground were also released, with the 20B model improving from 64.71% to 74.79% on FinanceReasoning and running on a single 80GB GPU.
同时开放了 20B 金融模型权重和无认证的测试平台；该 20B 模型在 FinanceReasoning 上从 64.71% 提升至 74.79%，可在单块 80GB GPU 上运行。