🔍 Read the full analysis: Transforming 350M AI Model Outputs In 100 GRPO Steps For Better Data Structuring on ThorstenMeyerAI.com
TL;DR
Liquid AI has publicly shared a cost-effective method to fine-tune its 350-million-parameter model using Group Relative Policy Optimization (GRPO). The process improves the model’s ability to produce structured, schema-compliant outputs, raising scores on the IFStruct benchmark from 22.6% to 29.7% after only 100 training steps on minimal data, accessible on free-tier hardware.
Liquid AI has released a fully open-source, inexpensive method for fine-tuning its 350-million-parameter LFM2.5 model using Group Relative Policy Optimization (GRPO), resulting in a notable improvement in structured-output compliance. The approach boosts the model’s score on the IFStruct benchmark from 22.6% to 29.7% after roughly 100 training steps, using only about 500 samples. This process can be completed on free-tier hardware such as Colab or Kaggle GPUs, making advanced fine-tuning accessible to small developers and researchers.
The method involves applying GRPO fine-tuning with the TRL library on a GPU, paired with local evaluation on a MacBook Pro equipped with an Apple M1 Max and 36 GB of memory, using llama.cpp for serving the model through an OpenAI-compatible endpoint. The base model was initially evaluated locally with a BF16 GGUF build, achieving 22.6% accuracy on the 2,000-sample IFStruct test set, aligning with the original benchmark’s reported baseline of 21.1%. The training data was sourced from NVIDIA’s Nemotron RL dataset, which pairs prompts with JSON schemas and expected field counts, with data augmentation techniques applied to improve format adherence.
Specifically, 40% of prompts received added instructions to return outputs within fenced code blocks, encouraging the model to follow formatting instructions rather than always output raw JSON. The analysis revealed that the model struggled most with JSON format adherence, passing only 18% of JSON outputs, but performed better with YAML (27.2%) and wrapper-key outputs. Common errors included missing required fields, incorrect item counts, and type mismatches, with per-entity success rates varying from 45.8% for event ticket bookings to as low as 4.3% for recipes. The improvements demonstrate that small, resource-efficient runs can meaningfully enhance schema compliance in language models.
Impact of Cost-Effective Fine-Tuning on Small Models
This development matters because it shows that small language models can be significantly improved with minimal resources, lowering barriers for developers and researchers lacking large-scale infrastructure. The roughly seven percentage point increase in schema adherence, achieved through only 100 steps on publicly available hardware, suggests that targeted, task-specific fine-tuning can bring small models closer to the performance of larger systems in structured output tasks. This is particularly relevant for real-world applications where reliable, parseable outputs are essential for downstream integration, such as automated data extraction or form filling.
By making the process openly available and reproducible, Liquid AI advocates for democratizing AI development, enabling broader experimentation and deployment of small, efficient models that meet practical needs without costly training pipelines. While the results are promising, the authors caution that the generalizability beyond the IFStruct benchmark and stability across different serving stacks or quantizations remain to be tested, underscoring the need for further validation.
GPU cloud computing for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on Structured Output and Benchmarking
Structured output, such as JSON or YAML, is a common requirement in applications that rely on language models to generate data in specific formats. Historically, many benchmarks have focused on reasoning or extraction, with less emphasis on measuring a model’s ability to produce valid, parseable structured data. The IFStruct benchmark, maintained by Liquid AI, provides a standardized way to evaluate a model’s adherence to schemas, serving as an important metric for practical deployment.
Prior to this work, larger models trained explicitly for structured tasks have shown high compliance, but small models often struggle, producing outputs with missing or malformed fields. Liquid AI’s recent effort demonstrates that even small models can be improved through targeted fine-tuning, using a lightweight, accessible process that requires minimal computational resources. This approach aligns with broader trends toward democratizing AI, reducing reliance on large-scale infrastructure.
“Our results show that a light fine-tuning process using GRPO can improve schema compliance scores significantly on a small model, even with minimal data and steps.”
— Liquid AI team
As an affiliate, we earn on qualifying purchases.
Limitations and Scope of the Results
It remains unclear how well these improvements will generalize to other structured-output tasks or different benchmarks beyond IFStruct. The reported gains are based on a specific setup with the LFM2.5-350M model, and stability across various serving stacks or quantization schemes has not yet been established. Additionally, the comparison to larger models is framed as a motivation rather than a direct benchmark, so claims of matching larger models’ performance are not empirically verified in this context. Further testing is needed to evaluate the robustness and scalability of this approach across diverse applications.
As an affiliate, we earn on qualifying purchases.
Future Directions for Small Model Fine-Tuning
Next steps include scaling the sample size and training steps to explore further improvements, testing the recipe on other small models and structured tasks, and evaluating stability across different deployment stacks and quantization methods. The open-source pipeline invites community experimentation, which could lead to broader adoption and refinement. Researchers may also investigate combining GRPO with other fine-tuning techniques or applying it to more complex structured data scenarios. Ultimately, the goal is to establish accessible, reliable methods for enhancing small models’ output quality in real-world applications.
As an affiliate, we earn on qualifying purchases.
Key Questions
Can this fine-tuning method be applied to other small models?
Yes, the recipe is designed to be generalizable and can be adapted to other small language models with similar architectures, provided they support fine-tuning with GRPO.
How accessible is this process for individual developers?
The entire pipeline is available on GitHub and can be run on free-tier hardware like Google Colab or Kaggle, making it feasible for individual researchers and small teams.
Will this method improve performance on other benchmarks?
While promising on IFStruct, further testing is needed to confirm its effectiveness on different structured output tasks and benchmarks. The current results demonstrate a proof of concept rather than a universal solution.
Does this approach require extensive data or training time?
No, the fine-tuning process uses roughly 500 samples and only 100 steps, making it quick and inexpensive compared to large-scale training pipelines.
What are the limitations of this approach?
It remains uncertain how well the improvements generalize beyond the specific setup, and stability across different deployment environments has not yet been confirmed.
Primary source: Hugging Face · via ThorstenMeyerAI.com
