aitraining 0.0.36

Advanced Machine Learning Training Platform - IN DEVELOPMENT

Train state-of-the-art ML models with minimal code

English | Portugues

📚 Full Documentation →

---

📖 Comprehensive Documentation Available

Visit docs.monostate.com for detailed guides, tutorials, API reference, and examples covering all features including LLM fine-tuning, PEFT/LoRA, DPO/ORPO training, hyperparameter sweeps, and more.

---

AITraining is an advanced machine learning training platform built on top of AutoTrain Advanced. It provides a streamlined interface for fine-tuning LLMs, vision models, and more.

Highlights

Automatic Dataset Conversion

Feed any dataset format and AITraining automatically detects and converts it. Supports 6 input formats with automatic detection:

Format	Detection	Example Columns
Alpaca	instruction/input/output	{"instruction": "...", "output": "..."}
ShareGPT	from/value pairs	{"conversations": [{"from": "human", ...}]}
Messages	role/content	{"messages": [{"role": "user", ...}]}
Q&A	question/answer variants	{"question": "...", "answer": "..."}
DPO	prompt/chosen/rejected	For preference training
Plain Text	Single text column	Raw text for pretraining

aitraining llm --train --auto-convert-dataset --chat-template gemma3 \
  --data-path tatsu-lab/alpaca --model google/gemma-3-270m-it

32 Chat Templates

Comprehensive template library with token-level weight control:

• Llama family: llama, llama-3, llama-3.1
• Gemma family: gemma, gemma-2, gemma-3, gemma-3n
• Others: mistral, qwen-2.5, phi-3, phi-4, chatml, alpaca, vicuna, zephyr

from autotrain.rendering import get_renderer, ChatFormat, RenderConfig

config = RenderConfig(format=ChatFormat.CHATML, only_assistant=True)
renderer = get_renderer('chatml', tokenizer, config)
encoded = renderer.build_supervised_example(conversation)
# Returns: {'input_ids', 'labels', 'token_weights', 'attention_mask'}

GRPO Training with Custom Environments

Train with Group Relative Policy Optimization using your own reward environment. The env provides prompts and scores multi-turn episodes — GRPO generates completions, scores them, and optimizes:

aitraining llm --train --trainer grpo \
  --model deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B \
  --rl-env-module my_envs.hotel_env \
  --rl-env-class HotelEnv \
  --rl-num-generations 4 \
  --rl-max-new-tokens 256

Your environment implements three methods:

class HotelEnv:
    def build_dataset(self, tokenizer) -> Dataset:
        """Return HF Dataset with 'prompt' column."""

    def score_episode(self, model, tokenizer, completion, case_idx) -> float:
        """Run multi-turn episode from completion, return 0.0-1.0 score."""

    def get_tools(self) -> list[dict]:
        """Return tool schemas for generation (optional)."""

Custom RL Environments (PPO)

Build custom reward functions for PPO training with three environment types:

# Text generation with custom reward
aitraining llm --train --trainer ppo \
  --rl-env-type text_generation \
  --rl-env-config '{"stop_sequences": ["</s>"]}' \
  --rl-reward-model-path ./reward_model

# Multi-objective rewards (correctness + formatting)
aitraining llm --train --trainer ppo \
  --rl-env-type multi_objective \
  --rl-env-config '{"reward_components": {"correctness": {"type": "keyword"}, "formatting": {"type": "length"}}}' \
  --rl-reward-weights '{"correctness": 1.0, "formatting": 0.1}'

Hyperparameter Sweeps

Automated optimization with Optuna, random search, or grid search:

from autotrain.utils import HyperparameterSweep, SweepConfig, ParameterRange

config = SweepConfig(
    backend="optuna",
    optimization_metric="eval_loss",
    optimization_mode="minimize",
    num_trials=20,
)

sweep = HyperparameterSweep(
    objective_function=train_model,
    config=config,
    parameters=[
        ParameterRange("learning_rate", "log_uniform", low=1e-5, high=1e-3),
        ParameterRange("batch_size", "categorical", choices=[4, 8, 16]),
    ]
)
result = sweep.run()
# Returns best_params, best_value, trial history

Enhanced Evaluation Metrics

8 metrics beyond loss, with callbacks for periodic evaluation:

Metric	Type	Use Case
Perplexity	Auto-computed	Language model quality
BLEU	Generation	Translation, summarization
ROUGE (1/2/L)	Generation	Summarization
BERTScore	Generation	Semantic similarity
METEOR	Generation	Translation
F1/Accuracy	Classification	Standard metrics
Exact Match	QA	Question answering

from autotrain.evaluation import Evaluator, EvaluationConfig, MetricType

config = EvaluationConfig(
    metrics=[MetricType.PERPLEXITY, MetricType.BLEU, MetricType.ROUGE, MetricType.BERTSCORE],
    save_predictions=True,
)
evaluator = Evaluator(model, tokenizer, config)
result = evaluator.evaluate(dataset)

Auto LoRA Merge

After PEFT training, automatically merge adapters and save deployment-ready models:

# Default: merges adapters into full model
aitraining llm --train --peft --model meta-llama/Llama-3.2-1B

# Keep adapters separate (smaller files)
aitraining llm --train --peft --no-merge-adapter --model meta-llama/Llama-3.2-1B

---

Screenshots

Built-in chat interface for testing trained models with conversation history

Terminal UI with real-time W&B LEET metrics visualization

---

Installation

pip install aitraining

Requirements: Python >= 3.10, PyTorch

Quick Start

Interactive Wizard

aitraining

The wizard guides you through:

1. Trainer type selection (LLM, vision, NLP, tabular)
2. Model selection with curated catalogs from HuggingFace
3. Dataset configuration with auto-format detection
4. Advanced parameters (PEFT, quantization, sweeps)

Config File

aitraining --config config.yaml

Python API

from autotrain.trainers.clm import train
from autotrain.trainers.clm.params import LLMTrainingParams

config = LLMTrainingParams(
    model="meta-llama/Llama-3.2-1B",
    data_path="your-dataset",
    trainer="sft",
    epochs=3,
    batch_size=4,
    lr=2e-5,
    peft=True,
    auto_convert_dataset=True,
    chat_template="llama3",
)

train(config)

---

Comparison

AITraining vs AutoTrain vs Tinker

Feature	AutoTrain	AITraining	Tinker
Trainers
SFT/DPO/ORPO	Yes	Yes	Yes
PPO (RLHF)	Basic	Enhanced (TRL)	Advanced
GRPO	No	Yes (TRL 0.28)	Custom
Reward Modeling	Yes	Yes	No
Knowledge Distillation	No	Yes (KL + CE loss)	Yes (text-only)
Data
Auto Format Detection	No	Yes (6 formats)	No
Chat Template Library	Basic	32 templates	5 templates
Runtime Column Mapping	No	Yes	No
Conversation Extension	No	Yes	No
Training
Hyperparameter Sweeps	No	Yes (Optuna)	Manual
Custom RL Environments	No	Yes (3 types)	Yes
Multi-objective Rewards	No	Yes	Yes
Forward-Backward Pipeline	No	Yes	Yes
Async Off-Policy RL	No	No	Yes
Stream Minibatch	No	No	Yes
Evaluation
Metrics Beyond Loss	No	8 metrics	Manual
Periodic Eval Callbacks	No	Yes	Yes
Custom Metric Registration	No	Yes	No
Interface
Interactive CLI Wizard	No	Yes	No
TUI (Experimental)	No	Yes	No
W&B LEET Visualizer	No	Yes	Yes
Hardware
Apple Silicon (MPS)	Limited	Full	No
Quantization (int4/int8)	Yes	Yes	Unknown
Multi-GPU	Yes	Yes	Yes
Task Coverage
Vision Tasks	Yes	Yes	No
NLP Tasks	Yes	Yes	No
Tabular Tasks	Yes	Yes	No
Tool Use Environments	No	Yes (GRPO)	Yes
Multiplayer RL	No	No	Yes

---

Supported Tasks

Task	Trainers	Status
LLM Fine-tuning	SFT, DPO, ORPO, PPO, GRPO, Reward, Distillation	Stable
Text Classification	Single/Multi-label	Stable
Token Classification	NER, POS tagging	Stable
Sequence-to-Sequence	Translation, Summarization	Stable
Image Classification	Single/Multi-label	Stable
Object Detection	YOLO, DETR	Stable
VLM Training	Vision-Language Models	Beta
Tabular	XGBoost, sklearn	Stable
Sentence Transformers	Semantic similarity	Stable
Extractive QA	SQuAD format	Stable

---

Configuration Example

task: llm-sft
base_model: meta-llama/Llama-3.2-1B
project_name: my-finetune

data:
  path: your-dataset
  train_split: train
  auto_convert_dataset: true
  chat_template: llama3

params:
  epochs: 3
  batch_size: 4
  lr: 2e-5
  peft: true
  lora_r: 16
  lora_alpha: 32
  quantization: int4
  mixed_precision: bf16

# Optional: hyperparameter sweep
sweep:
  enabled: true
  backend: optuna
  n_trials: 10
  metric: eval_loss

---

Documentation

📚 docs.monostate.com — Complete documentation with tutorials, API reference, and examples.

Quick Links

• Getting Started
• LLM Fine-tuning Guide
• YAML Configuration
• Python SDK Reference
• Advanced Training (DPO/ORPO/PPO)
• Changelog

Local Docs

• Interactive Wizard Guide
• Dataset Formats & Conversion
• Trainer Reference
• Python API
• RL API Reference

---

License

Apache 2.0 - See LICENSE for details.

Based on AutoTrain Advanced by Hugging Face.

---

Monostate AI