| name | llm-dna-knowledge |
| description | Use when working on LLM model lineage, provenance, phylogenetic analysis, from-scratch verification, the project's verification targets (Solar-Open-100B, K-EXAONE, A.X-K1, HyperCLOVAX), reference families (Llama 3, Qwen 2.5, GLM, Mixtral, DeepSeek), or anything under experiments/llm-dna/. Routes through the project wiki/ before external search. |
llm-dna-knowledge
Project-local convention: always consult the curated wiki/ knowledge base (28 cross-linked pages) before answering or acting on LLM lineage and verification topics. The wiki is compiled from primary sources; rebuilding context from scratch each session wastes tokens and produces contradictions.
When to Use
- Questions about LLM-DNA, neighbor-joining, UPGMA, distance metrics, functional fingerprint
- Verification methodology: tokenizer / weight / architecture / behavior analysis
- The 3 dokpamo targets: Solar-Open-100B, K-EXAONE-236B, A.X-K1
- Reference clusters: Llama, Qwen, GLM, Mixtral, DeepSeek (in lineage context)
- Edits or scripts under
experiments/llm-dna/
- "from-scratch" controversy, Phase 1/2 evaluations, Stanford AI Index entries
Skip when the project has no wiki/ directory — fall back to external sources only.
Workflow
- Read
wiki/index.md first. It groups all 28 pages by type (concept / entity / methodology / infra / policy). Match the user's keywords to 3–5 candidate pages.
- 2-hop expansion via
[[wiki-link]]. Open each candidate, follow inline links one or two levels deep until you have full context. Example: "Solar from-scratch?" → solar-open-100b → [[from-scratch-debate]] + [[layernorm-fingerprint-fallacy]] + [[glm-family]].
- Cite using
[[page-title]]. Reserve plain markdown links for external sources only.
- Wiki not enough? Say "위키에 없음 — 외부 자료 보강" explicitly. If the new content has lasting value, drop it under
wiki/raw/YYYY-MM-DD_<source>.md and offer to ingest via documentation:llm-wiki.
Quick Reference
| Need | Page(s) |
|---|
| LLM-DNA core | llm-dna-overview, dna-extraction-pipeline, random-projection, inheritance-and-determinism |
| Tree construction | phylogenetic-tree, neighbor-joining, upgma, distance-metrics |
| Target models | solar-open-100b, k-exaone-236b, ax-k1 |
| Reference families | llama-3-family, qwen-25-family, glm-family, mixtral-8x7b, deepseek-v25 |
| Methodology | tokenizer-analysis, weight-analysis, architecture-analysis, behavior-analysis, model-provenance-testing, layernorm-fingerprint-fallacy |
| Infra | sagemaker-spot-training, aws-cdk-typescript, huggingface-hub-usage |
| Policy / context | dokpamo-project, from-scratch-debate |
Q&A vs Wiki
The project runs both docs/tutorial/ (chronological Q&A learning log) and wiki/ (compiled best understanding). Flow on a new question: wiki check → answer → tutorial Q&A append → wiki update if non-trivial. Drives both forward without duplication.
Content vs Operational — sister skill llm-dna-extraction-playbook
This skill is for content questions (what is X, where is Y in the wiki, how does NJ tree work). For operational work — submitting/debugging SageMaker training jobs, modifying dna_train.py/submit_dna.py/analyze.py, planning Phase 5/6 retries, cost analysis — defer to llm-dna-extraction-playbook. That sister skill encodes 9 architectural lessons from Phase 4–6 ($34 learning curve cost), CLAUDE.md "인프라 운영 규칙" 9 rules, and entrypoint patches required for correct llm-dna 0.2.x behavior. Do not answer "should I run job X with config Y" from this skill alone — route through the playbook.
Wiki Maintenance Triggers
Offer the user a wiki update when:
- A verification result is finalized (e.g., LLM-DNA tree analysis produces a conclusion)
- A new reference model joins the lineup
- A policy/definition shifts (e.g., "from-scratch" criteria)
- An external source is genuinely worth keeping (drop into
wiki/raw/, then llm-wiki ingest)
Common Mistakes
- Searching externally before reading
wiki/index.md — ignores 28 pages of compiled knowledge
- Reading one page only, skipping
[[link]] expansion — answers miss critical caveats (e.g., LayerNorm fingerprint fallacy)
- Editing
wiki/index.md by hand — overwritten by the next llm-wiki sync
- Modifying files in
wiki/raw/ — raw zone is immutable; new info goes in a new file
- Concluding a verification but not updating the wiki — next session repeats the external lookup
- Citing a page whose frontmatter says
confidence: low without flagging that uncertainty in the answer
