Typhoon-S-4B NitiBench-CCL Legal Agent (Research Preview)

Typhoon-S-4B NitiBench-CCL Legal Agent is a "Sovereign," domain-specific research artifact designed to demonstrate that domain-specific sovereignty—through InK-GRPO–based agentic RFT— can outperform brute-force scale through a carefully designed post-training strategy.

For more information, please read the full technical report on arXiv.

This checkpoint is not a general-purpose instruction model and not intended for production or real-world legal use.

Not a product model
Not a general Thai legal assistant
Not safe or reliable for legal advice
Not expected to be useful outside the intended evaluation setup If you are looking for a generally capable assistant or a model for real-world legal workflows, do not use this.

Training Overview

Typhoon-S-4B NitiBench-CCL Legal Agent is post-trained using GRPO-based RFT with two key extensions.

Agentic RFT

The model is trained as a multi-step agent operating in a controlled RAG environment.
Available tools:
- search: semantic retrieval over a Thai legal corpus
- read: document-level content access
GRPO is applied over entire interaction trajectories, not single turns.
Rewards focus on final-answer correctness.

InK-GRPO (Injected Knowledge GRPO)

GRPO is augmented with a stochastic auxiliary next-token prediction objective on in-domain Thai legal text:

$\mathcal{L} = \mathcal{L}_{\text{GRPO}} + \lambda \, b \, \mathcal{L}_{\text{CE}}, \quad b \sim \text{Bernoulli}(\rho)$

This allows domain knowledge injection during reinforcement fine-tuning without collapsing into pure supervised learning.

Training Data (High-Level)

Training is centered on NitiBench (CCL) and aligned Thai legal corpora:

RFT (GRPO):
- Question–answer tasks derived from airesearch/WangchanX-Legal-ThaiCCL-RAG (CCL split)
- Rewarded based on correctness against reference answers (LLM-as-a-judge in the paper setup)
CE data (InK-GRPO):
- In-domain Thai legal text for auxiliary next-token prediction

Exact datasets, filtering, and preprocessing are described in the Typhoon-S technical report and NitiBench documentation.

Evaluation

Agent-Based Evaluation Required

This model is only meaningful when evaluated using the official agentic setup:

https://github.com/scb-10x/typhoon-s/tree/master/evaluation

Evaluating this checkpoint outside the specified agent + RAG environment will produce non-comparable and misleading results.

NitiBench (Thai Legal Reasoning, Agentic)

Model	NitiBench Accuracy
Qwen3-4B-Instruct-2507 + Agent	46.11%
GPT-5 + Built-in Search	38.07%
GPT-5 + Agent	75.34%
GRPO + Agent	73.73%
Typhoon-S-4B NitiBench-CCL InK-GRPO + Agent	78.02%

Results are benchmark- and environment-specific and should not be interpreted as general legal competence.

How to Use (Research Only)

This checkpoint is intended only for:

Studying Agentic RFT and InK-GRPO behavior

Recommended Usage

Run the official agentic evaluation pipeline:

https://github.com/scb-10x/typhoon-s/tree/master/evaluation

Expected conditions:

Agent-style prompting
search / read tools enabled
Thai legal corpus aligned with NitiBench
Evaluation protocol unchanged

Intended Uses & Limitations

Intended Use

Research-only experimentation
Benchmark comparison under NitiBench agentic evaluation

Limitations & Risks

Not a deployable model
Not legal advice
Benchmark-specific optimization
Environment-dependent performance
No safety, bias, or robustness guarantees
May hallucinate statutes, cases, or interpretations

Citation

If you use this model or its methods, please cite the Typhoon-S technical report.
If you use the dataset, please cite NitiBench directly.

@misc{pipatanakul2026typhoonsminimalopenposttraining,
      title={Typhoon-S: Minimal Open Post-Training for Sovereign Large Language Models}, 
      author={Kunat Pipatanakul and Pittawat Taveekitworachai},
      year={2026},
      eprint={2601.18129},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.18129}, 
}