💬 NLP

Typhoon-S: Minimal Open Post-Training for Sovereign Large Language Models

The paper introduces Typhoon S, a minimal and open post-training recipe combining supervised fine-tuning, on-policy distillation, and small-scale RFT with InK-GRPO to demonstrate that high-quality, sovereign large language models capable of region-specific tasks like Thai legal reasoning can be developed under academic-scale resources without relying on massive instruction corpora or complex reinforcement learning pipelines.

Original authors: Kunat Pipatanakul, Pittawat Taveekitworachai

Published 2026-01-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kunat Pipatanakul, Pittawat Taveekitworachai

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brilliant, world-class student (a Large Language Model) who speaks perfect English and Chinese but struggles with the local dialect, culture, and laws of a specific country. Usually, to fix this, big tech companies throw massive amounts of money, supercomputers, and endless data at the problem. But what if a smaller team, like a university or a national government, wants to create their own "sovereign" AI that understands their local context without needing a billion-dollar budget?

This paper introduces Typhoon-S, a "minimalist recipe" for teaching these AI models to be both helpful general assistants and experts in local, high-stakes tasks (like Thai law), using very limited resources.

Here is how they did it, broken down into two main parts:

Part 1: Making the AI "Adoptable" (The General Assistant)

The Goal: Turn a raw, smart base model into a polite, helpful assistant that can follow instructions, write code, and chat naturally in the local language.

The Problem: If you just teach a model new instructions (Supervised Fine-Tuning or SFT), it often becomes "brittle." It's like a student who memorized the textbook answers but panics when the teacher asks a slightly different question or mixes up the language.

The Solution: The "Shadowing" Technique (On-Policy Distillation)
The authors used a two-step process:

The Lesson (SFT): They first gave the model a small, high-quality mix of English and Thai instruction data. This was like giving the student a crash course.
The Shadowing (OPD): Instead of just memorizing the answers, the model was asked to generate its own answers, and a "teacher" model (a much smarter AI) would watch and correct it in real-time.
- Analogy: Imagine a music student playing a song. In the old way, they just listened to a recording and tried to copy it. In this new way, the student plays, and a master musician sits next to them, whispering corrections while the student is playing. This helps the student learn how to recover from their own mistakes, making them much more robust and flexible.

The Result: They achieved this with just a few days of training on a small cluster of GPUs (about the size of a university lab). The resulting model, Typhoon-S-8B, became a strong general assistant that could chat, code, and switch between English and Thai smoothly, rivaling much larger models.

Part 2: Giving the AI "Sovereign Capability" (The Local Expert)

The Goal: Teach the model to handle complex, high-stakes local tasks, specifically Thai legal reasoning, where it needs to understand local laws and use tools to find answers.

The Problem: Standard AI training often just reinforces what the model already knows. If the model doesn't know a specific Thai law, standard training won't magically teach it that new fact. Also, standard training doesn't teach the AI how to use tools (like a search engine) to find information.

The Solution: The "Dual-Brain" Training (InK-GRPO)
The authors invented a new training method called Injected Knowledge GRPO (InK-GRPO).

The Brain Game: Imagine the AI is playing a game where it has to solve a legal puzzle.
- Brain A (The Player): It tries to solve the puzzle using a reward system (like a video game score) to get better at reasoning.
- Brain B (The Reader): While Brain A is playing, Brain B is quietly reading a stack of Thai legal documents.
- The Magic: The system randomly switches between these two modes. Sometimes the AI plays the game; sometimes it reads the documents. This allows the AI to learn new facts (the laws) while simultaneously learning how to use them to solve problems.

The Tool User (Agentic RFT):
They also taught the AI to use tools.

Analogy: Instead of trying to memorize every law in the world, the AI is taught to say, "I don't know that specific law, let me search the database," read the document, and then answer.
They trained the AI to do this in a loop: Think -> Search -> Read -> Think -> Answer.

The Result: The Typhoon-S-4B Legal Agent became incredibly good at Thai legal reasoning. Surprisingly, this small 4-billion-parameter model, trained with this specific method, actually outperformed a much larger, famous model (GPT-5) when both were given the same tools to search for answers.

The Big Picture

The paper proves that you don't need a massive supercomputer to build a powerful, sovereign AI.

For General Use: A simple mix of "lessons" and "shadowing" (SFT + OPD) is enough to make a model smart and helpful.
For Local Expertise: A clever mix of "playing games" and "reading textbooks" (InK-GRPO) allows a small model to learn new local facts and use tools effectively.

The Cost: They did all this with resources available to a typical university lab (a few days of training on 4 to 8 high-end GPUs), proving that smart design is more important than brute force scale for creating sovereign, localized AI.

Technical Summary: Typhoon-S: Minimal Open Post-Training for Sovereign Large Language Models

1. Problem Statement

The development of state-of-the-art Large Language Models (LLMs) is currently concentrated within a small number of organizations possessing massive compute resources and access to large-scale, high-resource language datasets (primarily English and Chinese). This creates a "resource gatekeeping" barrier for sovereign settings—scenarios where national or regional institutions must retain full control over model weights, training data, and methodologies while operating under strict transparency constraints and limited computational budgets.

Existing sovereign-adapted base models often exhibit strong regional knowledge but lag significantly in general instruction following, tool use, and agentic behaviors. The standard post-training recipe to bridge this gap—scaling massive general-purpose instruction corpora and complex Reinforcement Fine-Tuning (RFT) pipelines—is often inaccessible to smaller research groups or national initiatives due to engineering complexity and data requirements.

The paper identifies two core requirements for sovereign LLMs:

Adoptability: The ability to transform a base model into a general-purpose assistant capable of instruction following, math, code generation, and tool use.
Sovereign Capability: The ability to perform high-stakes, region-specific tasks (e.g., local legal reasoning, cultural knowledge) that are under-represented in general pretraining data.

The central research question is: Which post-training strategy can enable competitive performance with frontier models under academic-level resource constraints?

2. Methodology

The authors propose Typhoon S, a minimal and open post-training recipe designed to address the above requirements using Thai as a representative case study. The approach is divided into two distinct pipelines.

2.1 Pipeline for Adoptability (SFT + OPD)

To transform base models into instruction-tuned assistants, the authors utilize a two-stage pipeline:

Supervised Fine-Tuning (SFT):
- Data: A unified corpus combining open-source English instruction data (Tulu 3, Toucan Tool) with a dedicated Thai dataset. The Thai dataset is constructed via "AutoIF," translating real-user prompts (WildChat) and aggregating existing Thai instructions, augmented with random constraint translation and placement to improve code-switching robustness.
- Goal: Equip the base model with basic instruction-following and tool-use abilities.
On-Policy Distillation (OPD):
- Mechanism: Following SFT, the model undergoes Generalized Knowledge Distillation (GKD). Unlike offline distillation, OPD uses the student model to generate trajectories (on-policy data) which are then corrected by a teacher model.
- Implementation: The authors implement a lightweight framework using HuggingFace Transformers and TRL to enable full-logits distillation on limited hardware (4–8 H100 GPUs) via dynamic model swapping and CPU offloading.
- Objective: Minimize the divergence between the student and teacher token distributions to reduce train-inference mismatch and improve generalization.

2.2 Pipeline for Sovereign Capability (InK-GRPO)

To enhance domain-specific reasoning and knowledge injection, the authors introduce Injected Knowledge GRPO (InK-GRPO), an extension of the Group Relative Policy Optimization (GRPO) algorithm.

The Challenge: Standard RFT amplifies existing knowledge but struggles to inject new domain-specific facts or support complex agentic behaviors.
The Solution (InK-GRPO): The training objective is augmented with a stochastic cross-entropy (CE) loss computed on a separate in-domain text corpus (pretraining-style data).
- Formula: $L = L_{GRPO} + \lambda \cdot b \cdot L_{CE}$ , where $b \sim \text{Bernoulli}(\rho)$ .
- Mechanism: This allows the model to alternate between task-focused optimization (GRPO) and domain knowledge acquisition (CE loss) within the same training step.
Agentic RFT: The framework is extended to an agentic setting where the model interacts with a Retrieval-Augmented Generation (RAG) environment using search and read tools. The model learns to interleave reasoning with tool calls to solve multi-step problems.

3. Key Contributions

Definition of Sovereign Requirements: The paper formalizes the dual needs of adoptability (general assistant capabilities) and sovereign capability (region-specific high-stakes tasks) for resource-constrained settings.
Minimal Adoptability Recipe: A validated pipeline combining lightweight SFT and On-Policy Distillation (OPD) that transforms sovereign-adapted base models into competitive instruction-tuned assistants without massive scale.
InK-GRPO: A novel algorithm that augments GRPO with stochastic next-token prediction on in-domain data. This enables the simultaneous optimization of task performance and the injection of new domain knowledge, addressing a key limitation of standard RFT.
Agentic RFT Extension: Demonstrating that InK-GRPO can be effectively applied to multi-turn agentic reasoning scenarios with external tool use.

4. Results

The experiments were conducted using Thai as the primary test case, utilizing models ranging from 4B to 8B parameters trained on 4–8 H100 GPUs.

Adoptability Results

SFT vs. SFT+OPD: SFT alone was found insufficient for robust performance, particularly in code-switching and tool use. Adding OPD improved the average benchmark score by +6.49 points and significantly enhanced robustness in Thai code-switching tasks.
Full-Logits vs. Top-K: Full-logits distillation provided superior robustness in open-ended generation and multilingual code-switching compared to Top-K distillation, though Top-K was sufficient for constrained tasks.
Base Model Adaptation: The recipe successfully transformed the sovereignty-adapted base model ThaiLLM-8B into Typhoon-S-8B Instruct. This model surpassed the general-purpose Qwen3-8B on strict Thai benchmarks (Chat, Code-Switching, OpenThaiEval) while maintaining competitive performance on general English tasks.

Sovereign Capability Results

InK-GRPO Efficacy: On NitiBench (Thai legal reasoning), InK-GRPO achieved 19.30% accuracy compared to 15.82% for standard GRPO. On MIRAGE-Bench, it also showed consistent gains.
Data Source: Pretraining-style in-domain text was found more effective for the CE loss component than SFT-style instruction data, likely because it avoids over-constraining the policy.
Agentic Performance: In a multi-turn agentic RAG setup, Typhoon-S-4B Legal Agent (trained with Agentic InK-GRPO) achieved 78.02% accuracy on NitiBench, outperforming both GRPO-only training (73.73%) and GPT-5 with comparable agentic setups (75.34%).
General Capability Retention: Crucially, models trained with InK-GRPO showed no evidence of severe catastrophic forgetting. General capability scores remained stable (within a narrow range of 48.08%–49.55%) across all benchmarks, indicating that domain specialization did not degrade general reasoning or knowledge.

5. Significance and Claims

The paper argues that careful post-training design, rather than sheer scale, is the key to practical sovereign LLMs.

Resource Efficiency: The proposed recipes achieve high-quality results with academic-scale resources (approximately two days of 8-GPU training for adoptability and one day of 4-GPU training for sovereign capability).
Transparency and Control: By relying on open datasets and transparent, minimal pipelines, Typhoon S provides a viable path for national institutions to develop models they fully control, understand, and can adapt to local cultural and legal contexts.
Generalizability: While the study focuses on Thai, the authors claim the methodology is not tied to the specific dataset and is generalizable to other languages and cultural contexts.

The authors conclude that their approach offers a "practical path toward high-quality sovereign LLMs," demonstrating that complex, resource-heavy pipelines are not strictly necessary to achieve competitive performance in both general and region-specific domains.