# wdndev/tiny-llm-zh

从零实现一个小参数量中文大语言模型。

Repository: https://github.com/wdndev/tiny-llm-zh
Canonical: https://ross.abutalabs.com/products/tiny-llm-zh
Language: Python
License Family: other
Last push: 2024-08-22T15:06:32+00:00

## Health v2 (maintenance only)
Score: 25/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 0, release rhythm 35, longevity 65
- inputs: {"age_days": 913, "days_push": 741, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1078, forks 125 (observed 2026-08-28T04:03:29.856563+00:00)

## What it is
An educational project that implements a small-parameter Chinese large language model from scratch, covering the full pipeline: tokenizer training, pretraining, SFT, RLHF/DPO alignment, evaluation, quantization, and deployment. It uses a Llama-style architecture (RMSNorm, RoPE, MHA) with Transformers and DeepSpeed, and supports MoE, vLLM, and llama.cpp inference.

## Use cases
- learn how to train a large language model from scratch
- understand the full LLM pipeline from tokenizer to deployment
- train a small Chinese language model
- study RLHF and DPO alignment implementation
- experiment with MoE architecture in transformers
- practice pretraining and SFT with deepspeed
- learn LLM quantization and deployment with vLLM or llama.cpp

## When to choose
- you want a hands-on, end-to-end walkthrough of building and training an LLM
- you need a small, resource-friendly Chinese model for learning purposes
- you want to study modern LLM techniques like RoPE, MoE, DPO, and quantization in one codebase

## When to avoid
- you need a production-quality model with strong benchmark results
- you require a commercially licensed model (no license is specified)
- you need a large multilingual or English-focused model

## Facets
- artifact type: learning-resource
- maturity: active
- function: llm-training, llm-inference, machine-learning, deep-learning, rag
- domain: large-language-models, deep-learning, tutorials, machine-learning
- platform: python, cloud
- tags: chinese-llm, tokenizer-training, sft, rlhf, dpo, moe, deepspeed, vllm, llama-cpp, model-quantization, pretraining, educational-project, natural-language-processing, gpu, linux, docker

## Member repositories
- wdndev/tiny-llm-zh (main) score 25

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:29.856563+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:53:12.585444+00:00, confidence not recorded.
  - readme: https://github.com/wdndev/tiny-llm-zh (fetched 2026-08-28T04:03:29.856563+00:00, sha 2283e3b08f0d)
- Data as of 2026-08-30T08:39:29.467469+00:00.