RoseLLM: Learn the Whole Large Language Model Stack

RoseLLM is a textbook that can be executed and a codebase that can be studied. Its scope is the complete lifecycle of a large language model: data, modeling, pretraining, post-training, reinforcement learning, evaluation, serving, systems, agents, and safety.

Two ways to use this repository

Follow the curriculum

Start with the LLM learning roadmap. It orders the topics by dependency and attaches concrete mastery checks. It also identifies where the repository already provides a derivation, source lab, or primary-source reading path and where prerequisite textbook study is still required.

The first complete specialization is the Agentic Reinforcement Learning curriculum. It begins with probability and Markov decision processes, then reaches long-horizon agent training, distributed rollout systems, and source-level implementation. Read the long-horizon capability synthesis to separate what laboratories train into model weights from what context management, tools, memory, verification, and multi-agent runtimes add. Read the research standard before using the frontier-lab and open-industry evidence matrix.

For a quantitative orientation before entering the case studies, use the model training token ledger. It separates fresh pretraining, continued pretraining, post-training, corpus size, and context length across DeepSeek, Kimi, GLM, Qwen, other Chinese lineages, and international reference families. The companion DeepSeek release timeline distinguishes mainline checkpoints, specialist branches, API aliases, and temporary modes.

The Multimodal Foundation Models section is a second evidence-grounded specialization. It traces image, video, document, and audio representations through encoders, projectors, shared backbones, media decoders, and production cascades. Its Kimi and DeepSeek studies distinguish input understanding from media generation and technical disclosure from labels such as “native multimodal.”

The Coding-Agent Systems section separates models, agent harnesses, interaction surfaces, and hosted control planes. Its first source-level reconstruction follows both Reasonix Git roots and compares the current Go system with OpenCode, Pi Coding Agent, OpenAI Codex, and Claude Code across provider coupling, agent loops, persistence, extension boundaries, permissions, sandboxing, and core-source visibility.

The Instruction Following and Steerability chapter separates atomic constraints, multi-turn retention, instruction hierarchy, tool-policy compliance, structured decoding, and end-to-end product behavior. Its companion method map, vendor evidence audit, and production operations chapter span data construction, post-training, inference, agent controls, the disclosed practices of 27 leading developers, and a rule-level release system. The overview maps the major public benchmarks and preserves a dated GPT-versus-DeepSeek comparison.

Where a new reader should begin

Use this dependency order rather than jumping directly to the newest model report:

1. Orientation and prerequisites: this page, then the roadmap's five-pass loop and Levels 0–6. Use the annotated bibliography as a reading side rail.
2. Coding-agent architecture: systems map, then the Reasonix history and cross-project source comparison. Keep model capability separate from harness and product behavior.
3. Multimodal architecture: representation/fusion first, then Kimi and DeepSeek lineages, and finally image-generation architectures.
4. Model evaluation: instruction-following taxonomy, benchmark map, and private production-evaluation design. Keep checkpoint behavior separate from constrained decoding and agent-harness behavior.
5. Agentic RL foundations: curriculum map, terminology, long-horizon capability synthesis, history, mathematical foundations, step-by-step derivations, and algorithm families.
6. The actual training system: data and environments, end-to-end pipeline, inference prerequisites, rollout/training systems, evaluation and safety, then the source-level lab.
7. Research reconstruction: research standard, evidence matrix, then DeepSeek → GLM → Kimi → OpenAI/Anthropic/Google → Qwen/Meta/Mistral → open industry and community.

Keep the glossary open while reading. It is a lookup tool, not a final chapter.

Trace concepts into code

Use the implementation as a set of progressively more realistic laboratories:

Layer	Repository area	Questions it makes concrete
Transformer training	rosellm/rosetrainer/	How do attention, loss, checkpoints, tensor parallelism, and DDP fit together?
LLM inference	rosellm/roseinfer/	How do paged KV caches, prefix reuse, continuous batching, prefill/decode disaggregation, and streaming work?
GPU kernels	notebooks/cuda/	Where do memory traffic, tiling, fusion, and numerical precision determine performance?
Distributed primitives	notebooks/	What do all-reduce, reduce-scatter, bucketing, and sharding actually do?
Multimodal models	docs/multimodal/	How do media become model positions, when is integration “native,” and how can a language-centered model emit images or speech?
Model evaluation	docs/evaluation/	Which factuality, grounding, citation, calibration, tool-integrity, or instruction surface failed, what is the reference, and which guarantee came from the model or surrounding system?
Coding-agent systems	docs/coding-agents/	How do models, agent loops, tools, context, clients, permissions, and sandboxes combine into coding products?
Agentic RL	docs/agentic-rl/ and rosellm/roserlhf/	How does an interactive trajectory become an unbiased, stable policy update?

The standard of evidence

This knowledge base separates five different kinds of statements:

• Disclosed fact: explicitly stated by a model developer, paper, model card, repository, or dataset card.
• Confirmed artifact: directly observable in released weights, configuration, data manifests, or source code even when prose does not state it explicitly.
• Reproduced result: independently produced with a recorded environment, command, configuration, and artifact.
• Inference: a conclusion supported by disclosed facts but not itself disclosed. The assumptions must be written next to it.
• Unknown: information that is not public or cannot be verified.

This distinction is essential in frontier-model case studies. A public model architecture does not reveal the private data mixture; an API behavior does not prove the training recipe; and benchmark scores do not disclose a production agent pipeline. See the research standard for the full protocol.

Current focus

The knowledge base currently has four deep, source-cited focus areas:

• Model evaluation and reliability: factuality and grounding controls, vendor-disclosed practices, claim-level production operations, instruction-following training and control methods, 27-developer evidence audit, benchmark methodology, and application-specific acceptance suites;

• Coding-agent systems: runtime lineage, model/provider coupling, agent loops, tools, state, extensibility, permissions, sandboxing, and open-source boundaries, beginning with a Reasonix/OpenCode/Pi/Codex/Claude Code study;

• Multimodal foundation models: modality representation, fusion, native integration, understanding versus generation, and the Kimi/DeepSeek model lineages; and
• Agentic Reinforcement Learning (Agentic RL): reinforcement learning in which a Large Language Model (LLM) policy interacts with a stateful environment over multiple turns.

The Agentic RL treatment covers both the scientific object and the engineering system:

1. the Partially Observable Markov Decision Process (POMDP) and policy-gradient formulation;
2. prompts, observations, actions, tools, memory, and state transitions;
3. offline demonstrations, online rollouts, rejection sampling, and curriculum construction;
4. outcome, process, preference, safety, and cost rewards;
5. PPO, GRPO, RLOO, REINFORCE-style methods, and agent-specific credit assignment;
6. asynchronous generation, version control for policies, replay, staleness, fault-tolerant environments, and distributed optimization;
7. evaluation against contamination, judge bias, reward hacking, and non-deterministic environments; and
8. disclosed practices from DeepSeek, Zhipu GLM, Moonshot Kimi, and other labs.

The Model Evaluation, Factuality, and Reliability map adds a separate production path for unsupported generation. It distinguishes open-world factuality, source faithfulness, citation entailment, calibration, tool-result integrity, and instruction following; catalogs controls from data and post-training through retrieval, tools, verification, abstention, monitoring, and human review; and audits which parts leading vendors have actually disclosed.

Contribution standard

Contributions should read as self-contained technical writing for their intended audience. Citations should point to primary sources when available, and quantitative claims should identify the exact table, section, model version, and evaluation setting. Follow the documentation and mathematical rendering standard for Markdown, TeX, generated-site, and browser acceptance checks.