ML-BOM guide: Qwen-7B example in 0x24 (Considerations) mixes in Qwen3-specific facts and contradicts 0x22

オープン
  • 何か問題がありますか？
#89 コメント 1 件 リアクション 0 件 担当者 0 名 GitHub で見る

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
35/100
issue の種類
ドキュメント
明瞭さ
おおむね明確
活発さ
活発
技術スタック
huggingface, machine-learning
領域
documentation, machine-learning

調査の方向性

0x24-Design-Model-Card-Considerations.mdを0x22-Design-Model-Card-Parameters.mdと併せて読み、その後、引用されているQwen-7BおよびQwen3のソースを確認してください。Maintainerのモデル選択を解決し、影響を受ける例を一貫して更新し、bom-ref、モデルの事実情報、コンテキスト長に関する主張が互いに矛盾しないようにしてください。

索引モデルが issue の本文から書いたものです。

説明

t: ML-BOM

Summary

The guide's running example model is Qwen-7B (Alibaba, released 2023-08-03, bom-ref: pkg:huggingface/Qwen/Qwen-7B@ef3c5c9, used throughout the guide starting in 0x20-Design-Model-Component-Metadata.md). The considerations example content in 0x24-Design-Model-Card-Considerations.md attributes several Qwen3-specific (2025) facts, features, and terminology to this Qwen-7B example — producing both external factual errors and an internal self-contradiction against the guide's own hyperparameter example in 0x22.

Internal contradiction (within the guide itself)

  • 0x22:385-386 declares cdx:ai-ml:model:hyperparameter:context_length = "8192" for pkg:huggingface/Qwen/Qwen-7B@ef3c5c9.
  • 0x24:88 says, for the same bom-ref: "the model supports up to 131,072 tokens using YaRN scaling, its native pre-training context is limited to 32,768 tokens."

Two different context-length claims for the same declared component, within the same guide.

Facts that belong to Qwen3, not Qwen-7B

  • "Thinking Mode" (0x24:46, :87, :90, :181) — the thinking/non-thinking mode toggle was introduced with Qwen3; it doesn't exist in Qwen-7B (2023).
  • YaRN scaling to 131,072 / native 32,768 context (0x24:88, :125) — this is Qwen3's documented spec (see Qwen3 Technical Report). Qwen-7B's real context length is 8,192, per its own published config.json.
  • "32B or 235B models" and "Qwen-8B" as sibling models (0x24:123, :125) — these are Qwen3 sizes (dense up to 32B, MoE 30B-A3B/235B-A22B). The original Qwen (Qwen1) lineup was 1.8B/7B/14B/72B — no 32B, no 235B, no "Qwen-8B".
  • "trained on 36 trillion tokens across 119 languages," attributed to "Qwen-7B" by name (0x24:160) — this is Qwen3's documented pretraining corpus size/language coverage. Qwen-7B (2023) was trained on ~2.4 trillion tokens.
  • "Qwen3Guard" cited as the safety reference for the Qwen-7B example (0x24:148, :165) — Qwen3Guard (released September 2025) is a guardrail model family built on Qwen3, unrelated to Qwen-7B.
  • "more parallelizable MoE models" comparison (0x24:128) — Qwen-7B is dense with no MoE sibling in its own generation.
  • Typo "Qwen3 &B parameter model" (0x24:112) — literally names "Qwen3" instead of "Qwen-7B." Note Qwen3's dense lineup is 0.6B/1.7B/4B/8B/14B/32B — there is no Qwen3-7B, so this can't even be "corrected" to match a real Qwen3 size.

Why it matters

This is the guide's primary worked example for modelCard.considerations. A reader relying on it as a reference, or cross-checking it against the real Qwen-7B model card, will find internally inconsistent and externally incorrect data — undermining the guide's own point about ML-BOMs enabling reproducibility and verifiable claims (see 0x10-Introduction.md, "Reproducibility & Explainability").

Suggested fix — needs a maintainer call, hence an issue rather than a PR

Either:

  • (a) Rebase the whole considerations example (and the context_length hyperparameter in 0x22) on Qwen3-8B, since most of the existing prose already matches that model, or
  • (b) Rewrite the considerations content to reflect the real, historical Qwen-7B (2023) — dropping Thinking Mode / YaRN / Qwen3Guard / MoE references.

Verification sources

This issue was drafted with assistance from Claude Sonnet 5 (model id: claude-sonnet-5), which identified and fact-checked the inconsistency described above.

主要言語
CSS
スター
10
フォーク
13
30日以内にマージされた PR はありません

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

CycloneDX/guides のほかの issue

CycloneDX/guides の issue をすべて見る

似ている issue

Documentation の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。