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

[andife]

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

• https://huggingface.co/Qwen/Qwen-7B/blob/main/config.json
• https://github.com/QwenLM/Qwen/blob/main/README.md
• https://arxiv.org/pdf/2505.09388 (Qwen3 Technical Report)
• https://huggingface.co/Qwen/Qwen3-8B
• https://qwenlm.github.io/blog/qwen3/
• https://github.com/QwenLM/Qwen3Guard

---

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

[mrutkows]

This is an accurate assessment by the AI. However, the AI I used to help gather the info and the model author actually published this for Qwen3, but not for 7B as of authoring (less capable one at that point-in-time) could not locate a nice representative set of "considerations" for the Qwen 7B model and much more documented analysis was avail. for the 3 version and likely there is 80% overlap. My immediate reaction is to add a disclaimer saying "representative for LLMs" and then alot time to use more powerful AI (if I can get tokens) to attempt to create 7B-specific examples, but again the research/analysis papers must exist or the model publishers publish specific 7B limits as first-hand info. which they did for Qwen3.