Typhoon-S: Sovereign LLM Post-Training

• Typhoon-S is a minimal, open post-training pipeline that transforms base LLMs into efficient general-purpose and domain-specific assistants.
• It uses a three-stage process—supervised fine-tuning, on-policy distillation, and reinforcement fine-tuning—to ensure efficiency and transparency.
• The approach achieves competitive performance on multilingual benchmarks, excelling in low-resource, sovereign applications like Thai legal reasoning.

Typhoon-S is a minimal, open post-training recipe for transforming LLMs into both general-purpose assistants and specialists for high-stakes, region- or domain-specific applications under academic-scale compute and data constraints. It is explicitly designed to address the dual challenges faced in sovereign settings: the need to retain local control and transparency over model weights, training data, and deployment; and the requirement to operate under limited computing resources and transparent recipes, in stark contrast to frontier LLM pipelines reliant on massive instruction corpora, multi-stage reinforcement learning with human feedback, and industrial-scale GPU clusters (Pipatanakul et al., 26 Jan 2026).

1. Motivation and Core Definitions

The frontier of LLM development is characterized by centralized resource gatekeeping: most models are developed by a small number of organizations with privileged access to compute and data, predominantly in high-resource languages such as English and Chinese. Sovereign LLM efforts—undertaken by regional or national institutions—encounter two critical barriers:

• Compute/Data Constraints: Academic and public-sector entities generally lack access to large GPU clusters (e.g., 1,000+ H100s) and cannot amass instruction datasets on the order of hundreds of millions of examples.
• Transparency Requirements: Sovereign or public deployments demand open, auditable pipelines, including full visibility into all data and model updates.

Typhoon-S articulates two complementary objectives for post-training under these constraints:

• Adoptability: The rapid and efficient conversion of a base model (open-weight or sovereign-adapted) into a general-purpose instruction-following assistant capable of chat, math, code, tool use, and robust multilingual handling, using a data-efficient regimen and modest compute (≤8 GPUs, a few hundred thousand examples, ≈2 days).
• Sovereign Capability: The additional specialization of the above assistant to solve domain- or region-specific tasks (e.g., Thai legal reasoning) through the targeted injection of in-domain knowledge and agentic abilities, using small-scale reinforcement learning (RL) and auxiliary learning stages.

2. Typhoon-S Post-Training Workflow

Typhoon-S comprises a three-stage post-training pipeline, each designed to maximize efficiency and transparency:

A. Supervised Fine-Tuning (SFT)

This initial stage starts from a base model $p_\theta$ and minimizes the standard cross-entropy loss over a curated instruction–response corpus $\mathcal{D}_{\text{SFT}}$:

$$\mathcal{L}_{\text{SFT}} = -\,\mathbb{E}_{(x,y) \sim \mathcal{D}_{\text{SFT}}} \sum_{t=1}^{T} \log p_\theta(y_t \mid x, y_{<t})$$

The SFT set (≈340 K examples) balances:

• 200 K general-purpose instructions (Tulu 3)
• 100 K tool-use/agentic examples (Toucan Tool)
• 40 K Thai AutoIF-generated instruction–response pairs

B. On-Policy Distillation (OPD)

To bridge the train–inference distribution gap of conventional (offline) distillation, Typhoon-S uses Generalized Knowledge Distillation (GKD) on student model rollouts. At each step, with probability $\lambda=0.25$:

• Either sample $(x,y)$ from $\mathcal{D}_{\text{SFT}}$ (offline)
• Or generate $y \sim p_S(\cdot\mid x)$ then query the teacher $p_T$ for logits on that trajectory (on-policy)

The student then minimizes:

$$\mathcal{L}_{\text{KD}} = \mathbb{E}\left[\sum_{t=1}^{T} D_{\text{KL}}(p_T(\cdot \mid x, y_{<t}) \,\|\, p_S(\cdot \mid x, y_{<t}))\right]$$

Full-logits distillation is preferred for robustness on code-switching and complex tasks. A top-$K$ token approach is possible for lower compute but is less robust, particularly for language mixing.

C. Small-scale Reinforcement Fine-Tuning (RFT) with InK-GRPO

To inject factual and procedural knowledge not present in pretraining, Typhoon-S invokes compact RL fine-tuning, using the GRPO (Generalized Reward Policy Optimization, a PPO variant) objective. This is augmented with an auxiliary cross-entropy term on in-domain corpus (e.g., legal statutes):

$$\mathcal{L} = \mathcal{L}_{\text{GRPO}} + \lambda b \mathcal{L}_{\mathrm{CE}}$$

Where $b \sim \mathrm{Bernoulli}(\rho)$ controls randomization ($\rho=0.6$, $\lambda=0.1$). In agentic settings, the model interacts with “search” and “read” tools over a FAISS-indexed corpus, optimizing end-to-end for final-answer accuracy.

3. Thai Language Case Study

The Thai case exemplifies Typhoon-S in a live, low-resource, high-stakes context, leveraging both multilingual and domain-specific alignment:

Datasets

• English: Tulu 3 (200 K), Toucan Tool (100 K)
• Thai Instructions: Sourced from translated WildChat, WangchanThaiInstruct, Han, Typhoon Instruct
• Response Generation: Thai AutoIF, using large-scale teacher LLMs with code-verifiable criteria
• Augmentation: Constraint translation (EN ↔ TH), variation in prompt placement
• Domain Supervision: NitiBench-CCL and MIRAGE-Bench (Thai legal, used for RL and CE data)

Compute Footprint & Model Sizes

• Adoptability (8B): 8× H100, ≈2 days (SFT+OPD)
• Sovereign agent (4B): 4× H100, ≈1 day (agentic InK-GRPO)

Stage	Data Size	Compute	Core Hyperparameters
SFT	340K	8×H100, 2 days	AdamW, lr=2e-5, batch=32
OPD	160K	8×H100, fused w/ SFT	lr=1e-6, λ=0.25, 1 epoch
InK-GRPO (RFT)	160K	4×H100, 1 day	lr=1e-6, ρ=0.6, λ=0.1

4. Evaluation Protocols and Empirical Outcomes

Typhoon-S employs a full sweep of multilingual, agentic, and sovereign task benchmarks:

General Capabilities (EN+TH)

• MT-Bench EN/TH: LLM-as-judge, helpfulness/correctness
• IFEval EN/TH: Factual, verifiable constraints
• Code-switching robustness: Realistic TH–EN mixing
• Knowledge & Reasoning: GPQA (EN), MMLU Pro X (TH), OpenThaiEval (TH-native)
• Math: MATH500 EN/TH
• Code reasoning: LiveCodeBench
• Tool/agentic: BFCL (tool use), HotpotQA EN/TH (RAG)

Sovereign Benchmarks

• NitiBench (Thai legal QA accuracy)
• MIRAGE-Bench (TH legal domain)

Key Results

• SFT alone results in performance deficits and brittleness (avg 37.45, code-switching 65.4, agentic 0) compared to strong baselines (48.07, 96.2).
• Addition of OPD yields significant improvements (avg +6.5 pts to 43.94; code-switching to 93.4), recovers robust agentic behavior, and maintains base knowledge.
• Full-logits OPD dramatically outperforms top-$K$ (code-switching: 93.4 vs 69.8).
• Thai-specific data is vital: removing it results in ≈4 pt drop in SFT and impacts sovereign task performance after OPD.
• When applied to a sovereign-adapted base (ThaiLLM-8B), Typhoon-S (SFT+OPD) achieves superior performance (Thai avg 71.20, Qwen3-8B 66.66) and competitive overall scores (49.99 vs 54.02).
• InK-GRPO improves sovereign task accuracy on NitiBench (19.30% vs 15.82%) and MIRAGE (22.63% vs 20.99%), as well as agentic settings (NitiBench 78.02% vs GRPO 73.73% and GPT-5+Agent 75.34%).
• General-purpose performance remains stable across RFT variants (avg ≈48–49 pts), indicating no catastrophic forgetting.

5. Trade-offs, Limitations, and Implications

Typhoon-S attains competitive region-specific and general performance using a fraction of the data and compute of mainstream LLM regimes:

• Data/Compute Efficiency: 340K SFT + 160K OPD instructions (vs millions commonly used); two days on 8 GPUs (adoptability), one day on 4 GPUs (sovereign agent).
• Performance: Comparable or superior to state-of-the-art open-weight baselines on Thai-centric and general benchmarks.
• Transparency: All stages are open and auditable at the token and gradient level, supporting stringent sovereign oversight.

Limitations

• No exploration of pre-training or mid-training under resource constraints.
• Fixed hyperparameters (e.g., $\rho, \lambda$)—extensive ablations and tuning deferred.
• The Thai focus reflects available data and expertise; extension to additional low-resource environments remains an open direction.
• Long-term effects of repeated InK-GRPO (e.g., knowledge saturation, model drift) require further investigation.

A plausible implication is that Typhoon-S enables credible sovereign alternatives to closed LLMs without the need for massive training budgets, provided strong local data curation and staged post-training are feasible.

6. Prospects for Sovereign LLM Workflows

Typhoon-S provides a reproducible and minimal blueprint for regions or domains seeking to democratize advanced LLM capabilities under real-world resource limits. By decomposing LLM post-training into SFT, on-policy distillation, and compact agentic RFT/knowledge injection, it sidesteps reliance on proprietary data, closed tools, and large-scale infrastructure. Its demonstrated stability and efficiency suggest a practical path for the broader adoption of sovereign LLMs in diverse settings—pending future validation beyond the initial Thai case and further methodological refinements (Pipatanakul et al., 26 Jan 2026).