Cookbook - posttrainllm with Pydantic AI

Cookbook - posttrainllm with Pydantic AI

What you get:

• A posttrainllm specialist behind Pydantic AI’s OpenAI-compatible provider.
• A typed result model that validates the response.
• One runnable local demo.
• A structured-output benchmark template.
• A clear limit: this is not a replacement for broad reasoning models.

What this is not: automatic correctness. Pydantic validates shape; your eval still has to measure whether the values are right.

1. Pick The Specialist

Structured output works best when the model was trained on the schema family you care about. Function-calling and constrained-output data are the closest starting point:

export TINYGPT_MODEL=/tmp/posttrainllm-structured-specialist.tinygpt

2. Serve posttrainllm

posttrainllm serve "$TINYGPT_MODEL" --host 127.0.0.1 --port 8080

3. Configure Pydantic AI

Pydantic AI’s OpenAIProvider accepts a custom base_url, so posttrainllm can be used like any other OpenAI-compatible server:

from pydantic import BaseModel
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIChatModel
from pydantic_ai.providers.openai import OpenAIProvider

class TicketRoute(BaseModel):
    team: str
    priority: int
    summary: str

model = OpenAIChatModel(
    "posttrainllm",
    provider=OpenAIProvider(
        base_url="http://127.0.0.1:8080/v1",
        api_key="not-needed",
    ),
)

agent = Agent(model, output_type=TicketRoute)
result = agent.run_sync(
    "Route this ticket: customer cannot log in after SSO migration."
)
print(result.output)

Runnable example:

examples/pydantic-ai-posttrainllm/run.sh

4. Benchmark

Measure schema compliance and semantic accuracy separately:

posttrainllm run-bench \
  --model "$TINYGPT_MODEL" \
  --tasks json-schema-routing \
  --limit 50 \
  --out docs/artifacts/pydantic-ai-specialist-bench.jsonl

Record:

Honest Limitations

• Pydantic AI can validate the object, but it cannot make a weak model smart.
• Some OpenAI-compatible servers need profile tweaks for strict schemas; keep posttrainllm prompts simple until the constrained-decoding path is wired into serve.
• Do not publish the benchmark table until it contains real numbers.

See also: smolagents and personal code specialist.