🔍 Read the full analysis: A Practical Approach To Refining 350M AI Models For Clearer Output Structuring on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Liquid AI has publicly shared an inexpensive fine-tuning method for its 350-million-parameter LFM2.5 model using Group Relative Policy Optimization (GRPO). This approach improves the model’s ability to produce structured, schema-compliant outputs, achieving a nearly 7-point accuracy increase on the IFStruct benchmark with minimal training data and runtime, accessible on free-tier hardware.
Liquid AI has released a fully public, low-cost recipe for fine-tuning its 350-million-parameter LFM2.5 model using Group Relative Policy Optimization (GRPO), resulting in a significant improvement in structured output accuracy. This development makes schema-compliant AI outputs more accessible for developers working with small models, especially on free-tier hardware. The approach is detailed in a GitHub guide and demonstrates a roughly 7-percentage-point increase on the IFStruct benchmark, from 22.6% to 29.7%, after just 500 training samples and 100 steps. For more details, see the original analysis on fine-tuning techniques.
The fine-tuning process utilizes a small dataset derived from NVIDIA’s Nemotron-RL-instruction_following-structured_outputs, with data augmentation techniques to improve format adherence. This method aligns with recent advances in efficient fine-tuning approaches. The model’s baseline performance was evaluated locally using a BF16 GGUF build of LFM2.5-350M, passing 452 out of 2,000 samples (22.6%) on the IFStruct benchmark. After applying the GRPO fine-tuning, the score increased to 29.7%, with the process completed on hardware such as a free-tier Colab or Kaggle GPU, making it highly accessible.
The training involved approximately 500 samples and 100 steps, focusing on prompts augmented with instructions to return outputs inside fenced code blocks, encouraging the model to follow formatting directives. For an in-depth explanation, see the detailed guide on this page. Error analysis shows most failures involved missing required fields, wrong item counts, or type mismatches, with the most common issues occurring in JSON format compliance. The approach emphasizes small, task-specific improvements rather than large-scale model retraining, aiming to demonstrate that small models can achieve performance levels comparable to larger models with minimal resources.
Impact of Cost-Effective Fine-Tuning on Small Models
This development is significant because it shows that small, affordable models can be improved to produce more reliable, schema-compliant outputs with minimal training data and computational resources. The open-source recipe lowers barriers for developers who lack access to large-scale training infrastructure, enabling broader experimentation and deployment of structured-output AI applications. The roughly 7-point gain on the IFStruct benchmark, achieved with just 100 steps on free hardware, underscores the potential for accessible AI refinement techniques to enhance practical usability in real-world systems.
GPU cloud computing for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on Schema Compliance and Small Model Optimization
Structured output tasks, such as JSON or YAML formatting, are common in real-world applications of language models, but many models struggle to consistently produce valid, parseable outputs. Prior benchmarks often evaluate these capabilities as part of broader reasoning tasks, rather than as standalone metrics. Liquid AI’s recent efforts focus on demonstrating that small models, which are easier to deploy and fine-tune, can be improved to meet practical standards for schema adherence. The IFStruct benchmark, an open-source test for output validity, serves as a key measure in this effort. Historically, larger models have shown stronger performance, but their resource requirements limit accessibility for many users. This context underscores the importance of low-cost, small-model improvements, particularly for developers without access to extensive computational infrastructure.
“Our goal was to show that task-specific fine-tuning can significantly improve small model performance at minimal cost.”
— Thorsten Meyer, Liquid AI
As an affiliate, we earn on qualifying purchases.
Uncertainties About Generalization and Broader Applicability
It remains unclear how well the observed improvements will generalize to other structured-output tasks or different models. The reported gains are based on the IFStruct benchmark, which, while representative, does not guarantee similar results across varied datasets or real-world applications. Additionally, the stability of these improvements under different serving stacks, quantization schemes, or larger fine-tuning runs has not yet been established. The claim that small models can match larger ones in performance is presented as a motivation rather than a tested, benchmarked fact, and further validation is needed to confirm its broader relevance.
JSON formatting tools for AI outputs
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Reproducibility and Expansion
Developers and researchers are encouraged to reproduce the fine-tuning process using the provided GitHub notebook, testing the recipe on other small models or datasets. Future work may include scaling the sample size and number of steps, experimenting with different quantization techniques, and applying the GRPO approach to other structured-output tasks. Additional validation across diverse benchmarks and real-world scenarios will help determine the method’s robustness and practical impact. Liquid AI has indicated plans to explore these avenues, but specific timelines or results have not yet been announced.
AI model training on free-tier hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Can this fine-tuning method be applied to larger models?
While the current demonstration focuses on a 350M parameter model, the principles of GRPO fine-tuning could potentially be adapted for larger models. However, the resource requirements and expected gains may vary, and further testing is needed to confirm effectiveness at scale.
Is the fine-tuning process suitable for production deployment?
The process is designed to be lightweight and accessible, making it suitable for experimental and development purposes. For production, additional validation, stability testing, and possibly larger-scale fine-tuning would be recommended.
What are the main limitations of this approach?
The primary limitations include uncertain generalization beyond the IFStruct benchmark, potential variability across different serving stacks, and the fact that improvements are demonstrated on a small dataset with minimal steps. More comprehensive testing is needed for broader claims.
How can I reproduce this fine-tuning process?
The full pipeline is available on GitHub, including a notebook for fine-tuning on free-tier hardware like Colab or Kaggle. The guide details steps for local evaluation using llama.cpp and the IFStruct dataset.
Does this approach improve other aspects of model performance?
The focus of this work is on schema compliance and structured output accuracy. Its impact on other capabilities such as reasoning or general knowledge remains to be explored.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
