How eugeneyan/open-llms works
1. Repository shape
- This is not an application or library repo; it is a documentation/catalog repository.
- The entire “system” consists of a README.md plus a LICENSE, with the README serving as the primary product artifact.
2. Primary purpose
- The repo maintains a curated registry of open large language models that are claimed to be commercially usable.
- It also extends the catalog to code models, pre-training datasets, instruction-tuning datasets, alignment datasets, and evaluation resources.
3. Core information model
- The README is organized as several large tables.
- Each row acts like a record with consistent fields such as:
- model/dataset name
- release date
- checkpoint links
- paper/blog link
- parameter count or sample/token count
- context length
- license
- optional demo/“try it” link
- In architecture-diagram terms, the repo is a metadata index rather than a computation pipeline.
4. Main content subsystems
- Open LLMs: general-purpose language models, grouped in a single table.
- Open LLMs for code: code-focused models with the same metadata pattern.
- Pre-training datasets: datasets used to build base models.
- Instruction-tuning datasets: datasets for supervised chat/instruction finetuning.
- Alignment-tuning datasets: datasets for post-training alignment.
- Evals: outbound references to third-party leaderboards and benchmark suites.
5. Data flow
- There is no runtime data flow inside the repo.
- The practical flow is:
1. maintainers collect model/dataset metadata from external sources,
2. normalize it into the README tables,
3. readers follow outbound links to Hugging Face, GitHub, papers, blogs, and demos.
- So the repo functions as a human-readable hub of outbound references.
6. Boundary with external systems
- Almost every meaningful artifact lives outside this repo:
- model checkpoints on Hugging Face or vendor sites
- papers/blogs on publisher or company sites
- demos on Hugging Face Spaces or project pages
- The repo does not host models, datasets, or evaluation results; it only indexes them.
7. Governance and curation boundary
- The repository’s “business logic” is editorial: deciding which models qualify as open, commercially usable, or relevant.
- License interpretation is part of the curation boundary, not an automated enforcement mechanism.
- The disclaimer explicitly states the repo is not legal advice, so legal interpretation remains outside the system.
8. Licensing as a key architectural concern
- License type is a first-class field in the catalog because commercial usability is the organizing principle.
- The README distinguishes permissive licenses like Apache 2.0/MIT from more constrained ones like LLaMA/Qwen/Gemma-style custom terms and OpenRAIL-M.
- This makes the repo a license-aware selection index for model choice.
9. Technology/runtime profile
- There is no application runtime, backend, frontend, or build pipeline visible in the repository contents provided.
- The only material technology is Markdown documentation with outbound web links.
- Any “execution” happens externally in the model providers’ ecosystems.
10. What is intentionally absent
- No source code, tests, packaging, CI, APIs, or data processing jobs are present in the provided tree.
- There is no persistence layer, compute layer, or service boundary to diagram inside the repo.
- The repo is effectively static content with periodic manual updates.
11. How to draw it
- Model the repo as a curated documentation hub in the center.
- From it, draw outbound edges to:
- model checkpoints/providers
- dataset repositories
- papers/blog posts
- demo spaces
- benchmark/leaderboard sites
- Do not draw internal service components; the “components” are the catalog sections themselves.
12. Operational meaning
- The maintainers’ job is editorial maintenance: add entries, keep context lengths/licenses current, and expand the catalog.
- The reader’s job is discovery and comparison, using the tables to evaluate models/datasets quickly.
13. Net architectural summary
- This repo is a static knowledge base for open LLM selection, centered on metadata curation and outbound references.
- Its architecture is defined by taxonomy, licensing metadata, and external linkage—not by software execution.
Main entry points: README.md README · LICENSE LICENSE
Components 16
In-repo docs
Static documentation artifacts that define the catalog.
- README · Primary curated index of open LLMs, datasets, and evaluations.
README.md - LICENSE · Repository license governing redistribution of the catalog content.
LICENSE - Open LLMs · General-purpose open language models with licensing and metadata.
- Code LLMs · Code-focused models listed with the same comparison fields.
- Pre-training · Datasets used for base model pre-training.
- Instruction-tuning · Supervised instruction and chat tuning datasets.
- Alignment-tuning · Post-training alignment and preference datasets.
- Evals · Outbound references to leaderboards and benchmark suites.
- Row fields · Shared record shape: name, dates, links, size, context, license, demo.
- License filter · Commercial usability and license interpretation as editorial criteria.
External ecosystem
Outside systems referenced by the catalog entries.
- Checkpoints · Model weights and dataset artifacts hosted externally, commonly on Hugging Face.
- Papers & blogs · Primary technical writeups and announcements linked from entries.
- Demos · Interactive try-it pages and hosted demos.
- Leaderboards · Third-party benchmark suites and evaluation rankings.
- Vendor sites · Model and dataset pages hosted by companies or project owners.
- Publishers · Academic or publication hosts for papers and benchmark descriptions.
Connections 21
- README → LICENSE: governed by · The catalog is published under the repository license.
- README → Row fields: uses · All tables follow a common metadata record shape.
- README → Open LLMs: contains · Main table of general-purpose open models.
- README → Code LLMs: contains · Table of code-oriented open models.
- README → Pre-training: contains · Catalog of pre-training datasets.
- README → Instruction-tuning: contains · Catalog of instruction-tuning datasets.
- README → Alignment-tuning: contains · Catalog of alignment-tuning datasets.
- README → Evals: contains · Index of external eval resources and leaderboards.
- README → License filter: curates by · License status is a first-class selection criterion.
- Open LLMs → Checkpoints: links to · Model rows point to external checkpoints and artifacts.
- Open LLMs → Papers & blogs: links to · Entries reference papers, blogs, and announcements.
- Open LLMs → Demos: links to · Optional try-it demos are linked when available.
- Open LLMs → Vendor sites: links to · Some models are hosted on vendor or project pages.
- Code LLMs → Checkpoints: links to · Code model artifacts live outside the repo.
- Pre-training → Checkpoints: links to · Dataset records reference hosted dataset repos.
- Instruction-tuning → Checkpoints: links to · Instruction datasets are linked externally.
- Alignment-tuning → Checkpoints: links to · Alignment datasets are linked externally.
- Evals → Leaderboards: points to · The eval section is an outbound index of benchmarks.
- Evals → Publishers: points to · Benchmark descriptions and papers are hosted externally.
- README → Papers & blogs: references · The README serves as a hub of curated outbound references.
- README → Checkpoints: references · Readers use the catalog to navigate to external artifacts.