ParaWriter-sft-stage1-4B

它能做什么 / What It Can Do

给它一篇论文，它可以告诉你：

快速体验 / Quick Demo

下面用一篇真实的 ICLR 2024 论文做演示。

输入：一篇完整论文

markdown
Copy code
论文标题：The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning

Abstract:
Alignment tuning has become the de facto standard practice for enabling base
large language models (LLMs) to serve as open-domain AI assistants. The alignment
tuning process typically involves instruction learning through supervised
fine-tuning (SFT) and preference tuning via reinforcement learning from human
feedback (RLHF). A recent study, LIMA, shows that using merely 1K examples for
SFT can achieve significant alignment performance, suggesting that the effect of
alignment tuning might be "superficial." This raises questions about how exactly
the alignment tuning transforms a base LLM.

We analyze the effect of alignment tuning by examining the token distribution
shift between base LLMs and their aligned counterparts. Our findings reveal that
base and aligned models perform nearly identically on 77.7% of token positions,
with the shifts occurring primarily on stylistic tokens. Based on these findings,
we introduce URIAL (Untuned LLMs with Restyled In-context ALignment), which
achieves effective alignment purely through in-context learning with as few as
three constant stylistic examples. Results demonstrate that base LLMs with URIAL
can match or even surpass the performance of SFT/RLHF models.

Introduction:
Base large language models (LLMs) that are only pre-trained on unsupervised text
corpora typically cannot directly serve as open-domain AI assistants. In order to
align these base LLMs as helpful and harmless assistants, recent research has
focused on alignment tuning through SFT and RLHF...

On the other hand, LIMA proposes the "Superficial Alignment Hypothesis," which
argues that alignment tuning might simply teach base LLMs to select a
subdistribution of data formats for interacting with users...

To this end, we investigate the effects of alignment tuning by directly comparing
token distributions between base LLMs and their aligned versions. Surprisingly,
we find that 77.7% tokens are unshifted, and shifted tokens are predominantly
stylistic (discourse markers, safety disclaimers)...

Based on our findings, we pose the research question: how effectively can we
align base LLMs without any tuning? We propose URIAL, which achieves alignment
through in-context learning with only three constant stylistic examples...

[后续正文包含 method、experiments 等]

任务示例 1：识别论文核心认知更新 / Central Realization

问模型：这篇论文让审稿人产生了什么核心认知更新？是"原来是这样"还是"原来还能这样"？

模型回答：

json
Copy code
{
  "realization_text": "Alignment tuning 并没有真正改变 base LLM 的知识，它只是在极少数 token 位置上做了风格调整（77.7% 的 token 完全不变）。因此不需要微调，仅靠 in-context learning 就能实现对齐。",
  "realization_type": "原来是这样",
  "why_this_type": "这篇论文揭示了 alignment tuning 的本质机制——表面看很复杂的 SFT+RLHF 流程，实际只改变了风格 token。这是对已知现象（alignment 有效）的深层解释，属于'原来是这样'类型。"
}

任务示例 2：分析 Introduction 段落结构 / Paragraph Role Analysis

问模型：这篇论文的 Introduction 每段分别承担什么功能？

模型回答：

json
Copy code
[
  {
    "paragraph_id": "intro_p1",
    "role": "scene_setting",
    "function": "建立研究背景：base LLM 需要通过 alignment tuning 才能用作 AI 助手，这是当前标准做法。"
  },
  {
    "paragraph_id": "intro_p2",
    "role": "gap_analysis",
    "function": "引入 LIMA 的'浅层对齐假说'，指出缺乏直接的定量证据，提出需要深入分析的动机。"
  },
  {
    "paragraph_id": "intro_p3",
    "role": "insight_delivery",
    "function": "报告 token distribution shift 分析的核心发现：77.7% token 不变，shifted 的主要是风格 token。"
  },
  {
    "paragraph_id": "intro_p4",
    "role": "method_preview",
    "function": "基于发现提出研究问题和 URIAL 方法：不微调，仅用 3 个 ICL 示例实现对齐。"
  },
  {
    "paragraph_id": "intro_p5",
    "role": "evidence_preview",
    "function": "预览评估方案和主要结果：URIAL 在多维度上达到甚至超过 SFT/RLHF 模型的表现。"
  },
  {
    "paragraph_id": "intro_p6",
    "role": "contribution_summary",
    "function": "总结三点贡献并强调对未来 alignment 研究方向的意义。"
  }
]

任务示例 3：检测承诺-兑现闭环 / Commitment-Evidence Mapping

问模型：Introduction 里做了哪些 claim？这些 claim 在后文实验中被兑现了吗？

模型回答：

json
Copy code
{
  "explicit_commitments": [
    {
      "claim": "Alignment tuning 主要改变风格 token，知识来自 base LLM 本身",
      "source_paragraph": "intro_p3",
      "fulfilled_in": "Section 2.2, Figure 2（token distribution shift 分析）",
      "fulfillment_strength": "strong"
    },
    {
      "claim": "URIAL 仅用 3 个常量 ICL 示例即可实现有效对齐",
      "source_paragraph": "intro_p4",
      "fulfilled_in": "Section 3.3, Table 1（多维度评估对比）",
      "fulfillment_strength": "strong"
    },
    {
      "claim": "Base LLM + URIAL 可以匹配甚至超过 SFT/RLHF 模型",
      "source_paragraph": "intro_p5",
      "fulfilled_in": "Table 1, Figure 1（Mistral-7b-Instruct、Llama-2-70b-chat 对比）",
      "fulfillment_strength": "strong"
    }
  ],
  "implicit_promises": [
    "方法具有跨模型泛化性（实际在 Llama-2、Mistral 上都测了）"
  ]
}

使用方法 / Usage

Python (Transformers)

python
Copy code
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "bhxdianzhang/ParaWriter-sft-stage1-4B"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)

# 把论文的 abstract + introduction + method + experiments 拼成输入
paper_content = """<PAPER_BODY>
... 你的论文正文 ...
</PAPER_BODY>"""

instruction = """You are an expert academic paper analyst. Given the body of a research paper,
identify the paper's CENTRAL REALIZATION — the single most important cognitive update it
delivers to its reader. Determine whether it is '原来是这样' or '原来还能这样'."""

messages = [{"role": "user", "content": f"{instruction}\n{paper_content}"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.3, do_sample=True)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

vLLM 部署

bash
Copy code
vllm serve bhxdianzhang/ParaWriter-sft-stage1-4B --trust-remote-code

python
Copy code
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")

response = client.chat.completions.create(
    model="bhxdianzhang/ParaWriter-sft-stage1-4B",
    messages=[{"role": "user", "content": "...论文内容 + 任务指令..."}],
    temperature=0.3,
)
print(response.choices[0].message.content)

训练详情 / Training Details

这是 Stage 1 / This is Stage 1

当前模型是第一阶段训练的结果，专注于理解和诊断——看懂论文的叙事结构。

第二阶段（Stage 2）将加入生成能力——根据论证结构直接生成 abstract/introduction 的段落和大纲。

This is the Stage 1 model, focused on understanding and diagnosis — comprehending paper narrative structure. Stage 2 will add generation capabilities — directly producing abstract/introduction paragraphs based on argument structure.

ParadoxGPT 模型家族 / Model Family

Citation

bibtex
Copy code
@misc{paradoxgpt2026,
  title={ParadoxGPT: Learning the Reviewer's "Aha Moment" for Better Scientific Writing},
  author={Heng Zhang},
  year={2026},
  url={https://huggingface.co/bhxdianzhang/ParaWriter-sft-stage1-4B}
}

License

Apache 2.0