Training Model Behavior in Agentic Systems

 

DeepFabric generates synthetic training data for language models and agent evaluations. By combining reasoning traces with tool-calling patterns, it creates high-quality, domain-specific datasets that teach models to think, plan, and act effectively, call tools correctly, and conform to strict schema structures.

What sets DeepFabric apart from other dataset generation tools is its ability to ensure high diversity yet domain-anchored relevance through unique topic graph generation algorithms. This guides sample creation to cover all necessary subtopics while avoiding redundancy, which is where other tools often fall short, resulting in model overfit.

Constrained decoding and response validation, along with real tool executions within isolated webassembly environments, ensure that generated samples strictly adhere to structured schema, variable constraints, and execution correctness, ensuring datasets have exact syntax and structure for use in model training pipelines. Tool definations can be either directly imported from MCP (Model Context Protocol) server schemas and automatically mocked, real life interfaces along with a standard set of common tools (list_files(), 'read_file() etc)

Once your dataset is generated, it can be automatically uploaded to Hugging Face and directly imported into popular training frameworks like TRL, Unsloth, and Axolotl.

Post-training, DeepFabric's built-in evaluation engine assesses model performance, whereby models prove their capabilities on unseen tasks derived from training splits—covering evaluation-only questions, answers, and tool traces.

Quickstart

DeepFabric can be used in several ways, as a library, CLI tool, or via YAML configuration. Here's a quick example using the CLI:

pip install deepfabric

export OPENAI_API_KEY="your-api-key"

deepfabric generate \
  --topic-prompt "Python programming fundamentals" \
  --generation-system-prompt "You are a Python expert" \
  --mode graph \
  --depth 3 \
  --degree 3 \
  --num-samples 9 \
  --batch-size 3 \
  --provider openai \
  --model gpt-4o \
  --output-save-as dataset.jsonl

This generates a topic graph and creates 27 unique nodes, then generates 27 training samples saved to dataset.jsonl, giving you 100% topic coverage.

Configuration

DeepFabric also uses YAML configuration with three main sections and optional shared LLM defaults

Note

The following uses mocked tool execution, so will require a runing Spin service, which we provide in a docker image:

docker run -d -p 3000:3000 ghcr.io/always-further/deepfabric/tools-sdk:latest`

Save the following as config.yaml:

# Optional: Shared LLM defaults (inherited by topics and generation)
llm:
  provider: "openai"
  model: "gpt-4o"
  temperature: 0.7

# TOPICS: Generate the topic tree/graph
topics:
  prompt: "Building production-ready REST APIs with Python"
  mode: tree                    # tree | graph
  depth: 3
  degree: 3
  save_as: "topics.jsonl"
  # Optional: Override shared LLM settings
  llm:
    model: "gpt-4o-mini"        # Use cheaper model for topics

# GENERATION: Create training samples from topics
generation:
  system_prompt: |
    You are an expert Python backend developer specializing in REST API design.
    Create practical, production-ready code examples with clear explanations.
    Include error handling, type hints, and follow PEP 8 conventions.
    Use the following tools to read, write, and list files in the virtual filesystem:
    - read_file
    - write_file
    - list_files

  # Additional instructions for sample generation
  instructions: |
    Focus on real-world scenarios developers encounter daily when building REST APIs with Python.
    Include both happy path and edge case handling.
    Provide context on when and why to use specific patterns or libraries.
    Ensure code is modular, testable, and maintainable.

  conversation:
    type: chain_of_thought      # basic | chain_of_thought
    reasoning_style: agent      # freetext | agent (for chain_of_thought)
    agent_mode: single_turn     # single_turn | multi_turn (for agent)
  
  # Tool configuration (required for agent modes)
  tools:
    spin_endpoint: "http://localhost:3000"  # Spin service for tool execution
    components:                 # Map component name to tool names
      builtin:                  # Routes to /vfs/execute
        - read_file
        - write_file
        - list_files
    max_per_query: 3            # Maximum tools per query
    max_agent_steps: 5          # Max ReAct reasoning iterations

  # Optional: Seed initial files into the spin before generation, used for tool calling
    scenario_seed:
      files:
        "Dockerfile": |
          FROM python:3.13
          WORKDIR /usr/local/app

          # Install the application dependencies
          COPY requirements.txt ./
          RUN pip install --no-cache-dir -r requirements.txt

          # Copy in the source code
          COPY src ./src
          EXPOSE 8080

          # Setup an app user so the container doesn't run as the root user
          RUN useradd app
          USER app

          CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8080"]
        "main.py": |
          def greet(name):
              return f"Hello, {name}!"

          if __name__ == "__main__":
              print(greet("World"))
        "config.json": |
          {
            "version": "1.0.0",
            "debug": true,
            "max_retries": 3
          }

  # Generation control and retry settings
  max_retries: 3                # Retries for failed generations
  sample_retries: 2             # Retries for validation failures
  max_tokens: 2000              # Max tokens per generation

  # Optional: Override shared LLM settings
  llm:
    temperature: 0.3            # Lower temp for consistent code

# OUTPUT: Final dataset configuration
output:
  # System prompt that goes INTO the training data
  # This is what the trained model will see as its system message
  system_prompt: |
    You are a helpful Python programming assistant specialized in REST API
    development. You provide clear, production-ready code with explanations.
    Always consider security, error handling, and best practices.

  include_system_message: true  # Whether to include system message in output
  num_samples: 4                 # Total training samples to generate
  batch_size: 3                 # Parallel generation batch size
  save_as: "api-dataset.jsonl"

 Optional: Upload to Hugging Face
 huggingface:
   repository: "your-username/api-dataset-training-name"
   tags: ["python", "programming"]

Run generation by sourcing the config.yaml:

deepfabric generate config.yaml

Generate, Train, Evaluate

DeepFabric returns standard HuggingFace datasets, making it easy to integrate with any training framework.

Colab Notebooks:

A quick way of seeing DeepFabric in action is via our notebooks in the notebooks/ folder or on Google Colab:

Qwen4b Blender MCP:

1. Generate Dataset

deepfabric generate config.yaml --output-save-as dataset.jsonl

Or upload to HuggingFace Hub:

deepfabric upload-hf dataset.jsonl --repo your-username/my-dataset

2. Load and Split for Training

from datasets import load_dataset
from transformers import AutoTokenizer

# Load from Hub
dataset = load_dataset("alwaysfurther/deepfabric-generic-tools", split="train")

# Split into train/eval
splits = dataset.train_test_split(test_size=0.1, seed=42)
train_ds = splits["train"]
eval_ds = splits["test"]

# Format using your tokenizer
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

def format_example(example):
    messages = [{k: v for k, v in msg.items() if v is not None}
                for msg in example["messages"]]
    return {"text": tokenizer.apply_chat_template(messages, tokenize=False)}

formatted_train = train_ds.map(format_example)

3. Train with TRL or Unsloth

from trl import SFTTrainer, SFTConfig

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=formatted_train,
    args=SFTConfig(output_dir="./output", num_train_epochs=3),
)
trainer.train()

4. Evaluate Your Model

from deepfabric.evaluation import Evaluator, EvaluatorConfig, InferenceConfig

config = EvaluatorConfig(
    inference_config=InferenceConfig(
        model_path="./output/checkpoint-final",  # Local path or HF Hub ID
        backend="transformers",
    ),
)

evaluator = Evaluator(config)
results = evaluator.evaluate(dataset=eval_ds)  # Pass HF Dataset directly

print(f"Tool Selection Accuracy: {results.metrics.tool_selection_accuracy:.2%}")
print(f"Parameter Accuracy: {results.metrics.parameter_accuracy:.2%}")
print(f"Overall Score: {results.metrics.overall_score:.2%}")

Evaluation

DeepFabric provides a comprehensive evaluation system to measure how well your fine-tuned models perform on tool-calling tasks.

Basic Evaluation

from datasets import load_dataset
from deepfabric.evaluation import Evaluator, EvaluatorConfig, InferenceConfig

# Load your evaluation dataset
dataset = load_dataset("your-username/your-dataset", split="test")

# Configure the evaluator
config = EvaluatorConfig(
    inference_config=InferenceConfig(
        model_path="./output/checkpoint-final",  # Local path or HF Hub ID
        backend="transformers",                   # "transformers" or "ollama"
        temperature=0.1,                          # Low temp for deterministic outputs
        max_tokens=2048,
    ),
    max_samples=100,           # Limit samples for quick testing (None for all)
    save_predictions=True,     # Save individual predictions
    output_path="eval_results.json",
)

# Run evaluation
evaluator = Evaluator(config)
results = evaluator.evaluate(dataset=dataset)

# Print summary
evaluator.print_summary(results.metrics)

# Cleanup GPU memory
evaluator.cleanup()

Evaluation with LoRA Adapters

from deepfabric.evaluation import Evaluator, EvaluatorConfig, InferenceConfig

config = EvaluatorConfig(
    inference_config=InferenceConfig(
        model_path="Qwen/Qwen2.5-7B-Instruct",    # Base model
        adapter_path="./output/lora-adapter",     # LoRA adapter path
        backend="transformers",
        load_in_4bit=True,     # 4-bit quantization
        max_seq_length=2048,
    ),
)

evaluator = Evaluator(config)
results = evaluator.evaluate(dataset=eval_dataset)

Understanding Evaluation Metrics

The evaluator computes several metrics for tool-calling tasks:

results = evaluator.evaluate(dataset=eval_dataset)
metrics = results.metrics

# Core metrics
print(f"Samples Evaluated: {metrics.samples_evaluated}")
print(f"Samples Processed: {metrics.samples_processed}")
print(f"Processing Errors: {metrics.processing_errors}")

# Tool-calling metrics
print(f"Tool Selection Accuracy: {metrics.tool_selection_accuracy:.2%}")
print(f"Parameter Accuracy: {metrics.parameter_accuracy:.2%}")
print(f"Execution Success Rate: {metrics.execution_success_rate:.2%}")
print(f"Response Quality: {metrics.response_quality:.2%}")
print(f"Overall Score: {metrics.overall_score:.2%}")

Metric	Description
tool_selection_accuracy	How often the model selects the correct tool
parameter_accuracy	How often tool parameters match expected values
execution_success_rate	Rate of valid, executable tool calls
response_quality	Quality score for non-tool responses
overall_score	Weighted combination of all metrics

Accessing Individual Predictions

results = evaluator.evaluate(dataset=eval_dataset)

# Iterate through individual sample evaluations
for pred in results.predictions:
    print(f"Sample {pred.sample_id}:")
    print(f"  Query: {pred.query}")
    print(f"  Expected Tool: {pred.expected_tool}")
    print(f"  Predicted Tool: {pred.predicted_tool}")
    print(f"  Tool Correct: {pred.tool_selection_correct}")
    print(f"  Params Correct: {pred.parameters_correct}")
    if pred.error:
        print(f"  Error: {pred.error}")

Evaluation from JSONL File

from deepfabric.evaluation import Evaluator, EvaluatorConfig, InferenceConfig

config = EvaluatorConfig(
    dataset_path="eval_dataset.jsonl",  # Load from file instead
    inference_config=InferenceConfig(
        model_path="./my-model",
        backend="transformers",
    ),
    output_path="results.json",
)

evaluator = Evaluator(config)
results = evaluator.evaluate()  # No dataset argument needed

Using Ollama Backend

from deepfabric.evaluation import Evaluator, EvaluatorConfig, InferenceConfig

config = EvaluatorConfig(
    inference_config=InferenceConfig(
        model_path="llama3.2:latest",  # Ollama model name
        backend="ollama",
        temperature=0.1,
    ),
)

evaluator = Evaluator(config)
results = evaluator.evaluate(dataset=eval_dataset)

Providers

Provider	Local/Cloud	Best For
OpenAI	Cloud	High quality, complex tasks
Anthropic	Cloud	Nuanced reasoning
Google Gemini	Cloud	Cost-effective at scale
Ollama	Local	Privacy, unlimited generation
OpenRouter	Cloud	Flexible model choice

Tool Tracing with Spin

DeepFabric supports real tool execution during dataset generation using the Spin Framework. Instead of simulating tool outputs, tools actually execute in isolated WebAssembly sandboxes, producing authentic training data.

Why Real Execution Matters

Traditional synthetic data generators simulate tool outputs, which creates unrealistic training data:

# Simulated (problematic)
Agent: read_file("config.json")
Result: {"setting": "value"}  # LLM hallucinated this content

With Spin integration, tools execute against real state:

# Real execution (accurate)
Agent: read_file("config.json")
Result: FileNotFound  # Actual filesystem state
Agent: write_file("config.json", "{...}")
Result: Written 42 bytes  # Real operation

ReAct-Style Execution

DeepFabric uses a ReAct (Reason-Act-Observe) loop for tool calling. The agent observes real results before deciding the next action:

Step 1: Agent thinks "I should check if config exists"
        -> Calls read_file("config.json")
        -> Observes: FileNotFound

Step 2: Agent thinks "Config doesn't exist, I'll create it"
        -> Calls write_file("config.json", content)
        -> Observes: Success

This produces training data where decisions are based on actual observations, not hallucinated assumptions.

Configuration

Enable tool tracing in your YAML config:

generation:
  conversation:
    type: chain_of_thought
    reasoning_style: agent
    agent_mode: single_turn

  tools:
    spin_endpoint: "http://localhost:3000"  # Spin service URL
    available:                              # Filter to specific tools
      - read_file
      - write_file
      - list_files
    max_agent_steps: 5                      # Max ReAct iterations

    # Optional: Seed initial state for scenarios
    scenario_seed:
      files:
        "config.json": '{"debug": true}'

Built-in VFS Tools

DeepFabric includes a virtual filesystem (VFS) component with these tools:

Tool	Description
read_file	Read content from a file
write_file	Write content to a file
list_files	List all files in the session
delete_file	Delete a file

Each session gets an isolated filesystem - changes don't persist between samples.

Running Spin Locally

cd tools-sdk
spin build
spin up

The Spin service runs at http://localhost:3000 by default.

Adding Custom Tools

You can extend DeepFabric with custom tools written in Python, JavaScript, Go, or Rust. See tool-traces.md for detailed documentation on:

• Creating custom Spin components
• Tool definition schemas
• Multi-language examples
• Containerization and deployment

Resources

• Documentation
• Examples
• Tool Tracing Guide
• Discord
• Issues

Development

git clone https://github.com/always-further/deepfabric
cd deepfabric
uv sync --all-extras
make test

Analytics

We collect anonymous usage metrics to improve DeepFabric. No personal data, prompts, or API keys are collected.

# Disable analytics
export ANONYMIZED_TELEMETRY=False