Llama 2:开放基础模型与对话微调模型Llama 2: Open Foundation and Fine-Tuned Chat Models
问题:开放权重模型能不能在有用性与安全性上接近闭源模型,同时把训练、微调与安全对齐的做法公开到可复用的程度?
证据位置:预训练在 §2(数据、40% 增量、双倍上下文、GQA);微调在 §3(SFT、拒绝采样、PPO);安全在 §4;观察与洞察在 §5,其中 §5.2 是 Limitations and Ethical Considerations;结论在 §7。
作者自陈局限:§5.2 列得很全——知识在预训练后停止更新;可能生成不实内容,例如不合格建议;有幻觉倾向;训练以英文为主,其他语言能力脆弱;可能生成有害、冒犯或偏见内容;安全微调有时过度保守,会出现过多拒答或堆砌安全说明;安全基准本身有局限,未必覆盖对抗输入、毒性内容与所有人群。摘要还说明 34B 模型因「没有足够时间做红队测试」而暂不发布。
它没主张的:没有说 Llama 2-Chat 在所有基准上超过闭源模型——原文是「may be a suitable substitute」,且基于自家人类评估;没有说安全对齐已经解决——§5.2 明说部分问题仍可能存在;也没有说多语言可用——原文说英文以外「fragile and should be used with caution」。Problem: can an open-weight model approach closed models on helpfulness and safety while publishing pretraining, fine-tuning and safety alignment in a reusable way?
Where the evidence sits: pretraining in Section 2 (data, 40% more tokens, doubled context, GQA); fine-tuning in Section 3 (SFT, rejection sampling, PPO); safety in Section 4; observations in Section 5, with 5.2 covering limitations and ethical considerations; conclusion in Section 7.
Stated limits: 5.2 is thorough — knowledge stops updating after pretraining; the model can generate non-factual content such as unqualified advice; it is prone to hallucination; training is predominantly English and other languages are fragile; it may produce harmful, offensive or biased content; safety tuning sometimes over-corrects into excessive refusal or too much safety boilerplate; and safety benchmarks themselves may miss adversarial inputs, toxic content and some demographic groups. The abstract also says the 34B model was withheld for lack of time to red-team it sufficiently.
What it does not claim: it does not claim Llama 2-Chat beats closed models on all benchmarks — the paper says it “may be a suitable substitute” based on its own human evaluations; it does not claim safety alignment is solved — 5.2 says some issues may remain; and it does not claim multilingual readiness, calling non-English performance fragile and to be used with caution.
关键参数Key specs
- 证据位置 / evidence
- §2 预训练;§3 微调;§4 安全;§5.2 局限与伦理;§7 结论
- 规模 / scale
- 7B、13B、70B 发布;34B 训练但未发布(红队时间不足)
- 关键数字 / key numbers
- 预训练语料 +40%;上下文长度翻倍;GQA
- 自陈局限 / stated limits
- 知识不更新、幻觉、英文为主、有害/偏见内容、安全过度保守(§5.2)
- 复现状态 / reproduction
- 未复现
官方原文怎么说What the source says
ranging in scale from 7 billion to 70 billion parameters
We also increased the size of the pretraining corpus by 40%, doubled the context length of the model, and adopted grouped-query attention
Llama 2-Chat is subject to the same well-recognized limitations of other LLMs, including a cessation of knowledge updates post-pretraining, potential for non-factual generation such as unqualified advice, and a propensity towards hallucinations.
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-17
- 许可 / 类型Licence / kind
- arxiv-preprint
- 许可说明Licence evidence
- https://arxiv.org/abs/2307.09288
- 类目Category
- 规模律Scaling laws
论文里的数字怎么读How to read the numbers
先分清测什么、怎么算、和谁比、有没有数据污染。拿不准就回原文看实验设置。Before citing a score, separate what it measures, how it is scored, what it compares against and whether contamination is controlled. If unsure, go back to the experiment setup in the paper. 看评测基准See benchmark cards · 看数据集卡See dataset cards