LLMsPracticalGuide: A Paper Index With a Licensing Appendix

What the repository is, and what it is not

The README opens by describing the project as a curated, still actively updated list of practical guide resources of LLMs, based on the survey paper Harnessing the Power of LLMs in Practice: A Survey on ChatGPT and Beyond. That sentence sets the scope: the deliverable is a bibliography with commentary, not a library. There is no install command in the material, no package manifest, no entry point, and no test suite. The primary language field is unknown and the licence field is unknown, which matters because the README itself says the project includes usage restrictions based on model and data licensing information. A repository that documents other projects' licences while not stating its own is a gap you should notice before you copy anything out of it.

The intended audience is practitioners who need orientation in NLP. The catalog is organised by concern rather than by chronology: models, data, tasks, efficiency, trustworthiness, alignment, prompting. Someone deciding which architecture family to read about first gets a map. Someone deciding which checkpoint to deploy gets nothing directly, because the repository points at papers, not at weights or serving code.

The model tree figure and its PowerPoint source

The most distinctive artifact here is an evolutionary tree of modern large language models, rendered as an image in the repository (imgs/tree.jpg) and described in the README as tracing the development of language models in recent years. The news section records that the authors plotted the figure in PowerPoint and released the source file source/figure_gif.pptx on 4/27/2023, then released source/figure_still.pptx on 4/29/2023 and replaced the repository figure with the still version.

That detail is more consequential than it looks. Because the editable source is shipped, you can correct or extend the tree for your own slides without redrawing it. It also tells you the figure's maintenance model: updates are manual edits to a slide deck, and the news entries show exactly which models were added when (AlexaTM, UniLM, UniLMv2 on 4/29/2023, along with a logo correction for Tk). A tree maintained this way drifts. The README does not state a schedule for revisiting it, and the news list stops in May 2023 even though the repository shows a later push date. Treat the figure as a snapshot of early 2023, not as a current model census.

How the catalog is organised and how you read it

The catalog splits into two model families and then into data and task concerns. Under Practical Guide for Models, BERT-style entries cover encoder-decoder and encoder-only work: BERT, RoBERTa, DistilBERT, ALBERT, UniLM, ELECTRA, T5, GLM, AlexaTM, ST-MoE. GPT-style entries cover decoder-only work: GPT, GPT-2, GPT-3, OPT, PaLM, BLOOM. Each line follows the same shape: a short name, the paper title, a year, and a link. There is no annotation field, no parameter count column, and no note on which checkpoints are downloadable.

Below models, the catalog branches into Pretraining data, Finetuning data and Test data/user data, then into Traditional NLU tasks, Generation tasks, Knowledge-intensive tasks, Abilities with Scaling, Specific tasks, Real-World Tasks, Efficiency, Trustworthiness, Benchmark Instruction Tuning, and Alignment. Alignment is the most subdivided branch, with Safety Alignment (Harmless), Truthfulness Alignment (Honest), Practical Guides for Prompting (Helpful), and Alignment Efforts of Open-source Community. The Harmless/Honest/Helpful framing is the structural idea worth keeping: it gives you a way to file alignment papers by objective rather than by venue. The trade-off is that a paper touching two objectives has to be filed under one, and the README does not say how that is resolved.

Usage and Restrictions: the section that carries the most weight

The news entries record two additions of substance: a usage and restrictions section, and a version covering commercial and research purposes, credited to Dr. Du on 5/8/2023. The README's own framing says the sources include usage restrictions based on the model and data licensing information. This is the part of the repository that is hardest to find elsewhere in one place, because licence terms for model weights and training corpora are scattered across model cards, dataset pages and custom research licences.

It is also the part most likely to age badly. Licence terms change, sometimes retroactively for new releases of the same model family. The README does not state when the restrictions table was last verified, and the section was added in May 2023. Using it as a first pointer is reasonable. Using it as the authority for a commercial deployment decision is not, because the repository is a secondary source describing primary licence text. The README also does not present itself as legal advice, and it should not be read as such.

Getting value out of it without installing anything

There is nothing to run. The practical workflow is: open the catalog, pick the branch that matches your problem, follow the paper links, and use the citation block when you reuse the material. The README supplies a BibTeX entry for the survey, keyed yang2023harnessing, with eprint 2304.13712 and primaryClass cs.CL. If you fork the tree figure into internal documentation, that citation is the condition the README attaches, phrased as a request rather than a licence term: use the resources, don't forget to cite our paper. Because the repository licence is not stated in the supplied material, you cannot confirm from this alone what redistribution of the figure is permitted. The homepage field points at the arXiv abstract page for version 2 of the paper, which is the canonical reference for the survey itself.

Pull requests are invited, specifically to refine the figure. That is the contribution channel the README names. It does not describe review criteria, a code of conduct, or how quickly submissions are handled.

Where it stops being the right tool

The repository cannot tell you whether a model actually performs on your task. It links to papers, and papers report results on their own benchmarks with their own prompts and decoding settings. If you need to compare two checkpoints on your data, this list gives you the reading, not the measurement. The catalog's Benchmark Instruction Tuning branch points at benchmark work, but the repository does not host a harness, a leaderboard, or evaluation scripts.

A second limit is currency. The news section's last dated entries are from 2023, and the model tree was frozen with the still version in April 2023. The repository metadata shows a push on 2026-04-08, so something changed, but the supplied material does not say what. You cannot infer from a push date that the model list was refreshed. If your question is about anything released after mid-2023, expect to leave this repository quickly.

A third limit is depth per entry. One line per model means no guidance on which of the BERT-style variants is worth the engineering cost for a given latency budget, and no signal on which GPT-style entries have usable open weights. The list is flat where a decision needs ranking.

How it differs from a hands-on LLM engineering guide

The README itself points elsewhere for build-oriented material, listing Building LLM applications for production by Chip Huyen and a Data-centric Artificial Intelligence repo, blog and paper. The contrast in approach is clean. The external guide addresses what to do when you put a model behind an API: caching, cost, evaluation, latency. This repository addresses what to read and what the licence allows. One is a set of engineering practices with code and trade-offs; the other is a bibliography with a licensing appendix and a diagram.

If your team is choosing between these, the question is whether the blocker is knowledge or execution. A team that cannot name the difference between encoder-only and decoder-only architectures gets more from the catalog's model branch than from a production blog post. A team that already knows the architecture families and is arguing about token costs gets nothing from a paper list. The README does not claim to serve the second team, and the absence of any code in the repository is consistent with that.

Maintenance cost and licence implications

Maintaining a fork of this repository is cheap in the mechanical sense and expensive in the editorial sense. Adding a line is trivial. Keeping the tree figure, the task taxonomy and the restrictions table mutually consistent is not, and the PowerPoint-based figure workflow means every structural change is a manual slide edit followed by an image export. The news log shows the authors doing exactly this in short succession in April and May 2023, which suggests the cost is real enough that updates cluster.

On licensing: the repository's own licence is not stated in the supplied material, so you cannot confirm the terms under which the list itself is redistributed. The README's citation request covers academic reuse of the work. The Usage and Restrictions section describes third-party model and data licences and is a secondary source for them. If you plan to ship a model that appears in that section, read the model's own licence text. Nothing here substitutes for that, and the repository does not present it as doing so.

Editorial conclusion

Adopt this repository as a citation index if you need to trace LLM papers back to the survey by Yang et al. and want a starting point on model and data usage restrictions. Do not adopt it if you need executable code, benchmark harnesses, or a maintained model comparison table, because the repository is a reading list whose last recorded push is 2026-04-08 and whose figure dates from 2023. Before using it, check the Usage and Restrictions section directly against each model's own licence, confirm whether the repository licence is stated anywhere in the files you are reading, and cite the arXiv paper 2304.13712 if you reuse the tree figure.

Official sources

1. Issues
2. Mooler0410/LLMsPracticalGuide on GitHub
3. Project website
4. README