Osmosis-Structure-0.6B: Small Language Model for Structured Outputs

Osmosis-Structure-0.6B is a specialized small language model (SLM) designed to excel at structured output generation. Despite its compact 0.6B parameter size, this model demonstrates remarkable performance on extracting structured information when paired with supported frameworks.

Our approach leverages structured output during training, forcing our model to only focus on the value for each key declared by the inference engine, which significantly improves the accuracy of the model’s ability to produce well-formatted, structured responses across various domains, particularly in mathematical reasoning and problem-solving tasks.

Results

We evaluate the effectiveness of osmosis-enhanced structured generation on challenging mathematical reasoning benchmarks. The following results demonstrate the dramatic performance improvements achieved through structured outputs with osmosis enhancement across different model families - the same technique that powers Osmosis-Structure-0.6B.

---

AIME 1983–2024 Performance

Structured Output vs. Structured w/ Osmosis

Model	Structured Output	Structured w/ Osmosis	Performance Gain
Claude 4 Sonnet	16.29%	62.59%	+284%
Claude 4 Opus	22.94%	65.06%	+184%
GPT-4.1	2.79%	39.66%	+1322%
OpenAI o3	92.05%	93.24%	+1.3%

---

DAPO-Math-17K Performance

Structured Output vs. Structured w/ Osmosis

Model	Structured Output	Structured w/ Osmosis	Performance Gain
Claude 4 Sonnet	15.52%	69.40%	+347%
Claude 4 Opus	15.28%	69.91%	+357%
GPT-4.1	10.53%	70.03%	+565%
OpenAI o3	91.14%	94.05%	+3.2%

---

Model Training

Osmosis-Structure-0.6B is built on top of Qwen3-0.6B. We first established a baseline format using 10 samples of randomly generated text and their JSON interpretations. We then applied reinforcement learning to approximately 500,000 examples of JSON-to-natural language pairs, consisting of either reasoning traces with their final outputs, or natural language reports with their expected structured formats.

We used verl as the framework to train our model and SGLang as the rollout backend. To enable structured training, we modified parts of the verl codebase to allow for per sample schema to be passed into the training data.

Usage

from ollama import chat
from pydantic import BaseModel

class Answer(BaseModel):
  answer: int

reasoning_trace = """
Problem: Solve for x in the equation 2x + 5 = 13

Let me work through this step by step:

First, I need to isolate the term with x. I'll subtract 5 from both sides:
2x + 5 - 5 = 13 - 5
2x = 8

Next, I'll divide both sides by 2 to solve for x:
2x ÷ 2 = 8 ÷ 2
x = 4

Let me verify this answer by substituting back into the original equation:
2(4) + 5 = 8 + 5 = 13 ✓

Ok, which means I got the correct answer, and I'm confident about my answer.
"""

response = chat(
  messages=[
    {
        "role": "system",
        "content": f"You are a helpful assistant that understands and translates text to JSON format according to the following schema. {Answer.model_json_schema()}"
    },
    {
      'role': 'user',
      'content': reasoning_trace,
    }
  ],
  model='Osmosis/Osmosis-Structure-0.6B',
  format=Answer.model_json_schema(),
)

answer = Answer.model_validate_json(response.message.content)
print(answer)