Home 零对冲(ZeroHedge)每日HackerNews

Logitloom：探索指令模型和基础模型的标记轨迹树
Logitloom: Explore token trajectory trees on instruct and base models

原始链接: https://github.com/vgel/logitloom

Logitloom 是一款用于探索语言模型令牌轨迹树（loom）的工具，允许你可视化和分析不同的可能输出序列。它可在 vgel.me/logitloom 获取。 要使用它，你需要一个支持助手预填充和对数概率的模型的 API 密钥。对于指令模型，推荐使用 Deepseek-v3（聊天类型，API 密钥来自 platform.deepseek.com）；对于基础模型，推荐使用 Hyperbolic 的 405-base（基础类型，API 密钥来自 app.hyperbolic.xyz）。将这些保存为预设以方便切换。 使用 Deepseek 聊天模型时，提供提示和可选的助手预填充，然后点击“运行”。使用 Hyperbolic 基础模型时，输入提示或预填充即可。Logitloom 将展开树，显示不同令牌序列的概率。 Logitloom 目前无需许可即可使用。开发中使用 Bun。注意：需要使用 Firefox 138 等较新的浏览器以避免 Bun 问题。

Hacker News new | past | comments | ask | show | jobs | submit login Logitloom: Explore token trajectory trees on instruct and base models (github.com/vgel) 9 points by kentricon 17 hours ago | hide | past | favorite | 1 comment gota 7 hours ago [–] How does "Currently unlicensed" work?I'd assume this (at least to the extent of whatever is published before a license is specified) be automatically assumed to be MIT license or something. Is that right?If a license that is more restrictive than that assumed default is chosen, can it be retroactively applied to the content that was published before it was chosen?None of these are rhetorical, I'm curious how this works reply Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact Search:
相关文章
  • 深seek证明器V2 2025-04-30
  • 深度搜索推理引擎开源之路 2025-04-14
  • DeepSeek-V3 技术报告 2025-03-27
  • Transformer 实验室 2025-04-14
  • DeepSeek Open Instra：5天内开源5个AI存储库 2025-02-22
    • 原文

    logitloom is a tool for exploring token trajectory trees (aka looming) on instruct and base models.

    Screenshot of logitloom

    Go to https://vgel.me/logitloom to use the deployed version of the tool.

    You will need a chat API that supports both assistant prefill (for the prefill setting and for expanding non-chosen branches), along with logprobs. I recommend deepseek-v3 for this. (NOT r1, which doesn't support logprobs via the Deepseek API.)

    • Base URL: https://api.deepseek.com/beta
    • API Key: Get one from https://platform.deepseek.com/api_keys. API keys are only stored and used locally, there's no backend.
    • Model: deepseek-chat
    • Type: chat

    (You can save this as a preset to easily switch back to later using Edit Presets.)

    You can now fill in a prompt, and optionally an assistant prefill, and click run to start expanding the tree.

    The best base model host currently is Hyperbolic's 405-base. I recommend using it directly, instead of via OpenRouter, which can introduce some issues. You will need a completions API that supports logprobs. (Most do.)

    • Base URL: https://api.hyperbolic.xyz/v1
    • API Key: Get one from https://app.hyperbolic.xyz/settings. API keys are only stored and used locally, there's no backend.
    • Model: meta-llama/Meta-Llama-3.1-405B
    • Type: base

    (You can save this as a preset to easily switch back to later using Edit Presets.)

    You can now fill in either a prompt or prefill, and click run to start expanding the tree. (Prompt and prefill are simply concatenated for base models, so use either one.)

    Currently unlicensed. TODO.

    Uses Bun for bundling and serving in development.

    • Serve dev: bun --hot index.html
      • Due to a Bun bug, you will need a recent browser, such as Firefox 138. oven-sh/bun#19469
    • Bundle (if you aren't thebes, you don't need this): ./build-for-website-and-copy.sh

    We currently vendor the OpenAI library due to some issues bundling it for browser with Bun. See vendor-openai.sh.

    联系我们 contact @ memedata.com