⚡ LitGPT (Custom Fork for RoBERTa MLM Pretraining)

This is a fork of lit-gpt, customized to support RoBERTa-base architecture for masked language modeling (MLM) pretraining. The implementation follows HuggingFace-style conventions while integrating seamlessly with the existing Lit-GPT codebase.

Features

• Roberta-base model architecture with MLM support
• Compatible with existing training/inference loops
• Test dataset for MLM pretraining
• Modular design for easy extension

Model Architecture

The Roberta-base implementation includes:

• 12 transformer layers
• 768 hidden dimensions
• 12 attention heads
• 3072 intermediate size
• 50,265 vocabulary size
• 512 maximum sequence length

Usage

Pretraining

To pretrain a Roberta-base model on the test dataset:

bash python -m litgpt pretrain \ --model_name roberta-base \ --data_dir path/to/test_data \ --out_dir path/to/output \ --train.global_batch_size 32 \ --train.micro_batch_size 4 \ --train.max_tokens 1000000 \ --train.learning_rate 1e-4 \ --train.lr_warmup_steps 2000 \ --train.save_interval 1000 \ --eval.interval 1000

Test Dataset

The test dataset is automatically created if it doesn't exist. It includes a small set of example sentences for testing the MLM pretraining pipeline. The dataset is created in the specified data directory with the following structure:

data_dir/ test.txt # Contains example sentences for MLM pretraining

Custom Dataset

To use your own dataset for MLM pretraining:

1. Create a new dataset class inheriting from TestMLMDataset
2. Implement the required methods:
   • _load_examples(): Load and tokenize your data
   • _create_masked_input(): Create masked input and labels
   • __getitem__(): Return inputids, labels, and attentionmask

Example:

```python from litgpt.data.test_data import TestMLMDataset

class CustomMLMDataset(TestMLMDataset): def loadexamples(self): # Load your custom data here examples = [] with open(self.datadir / "yourdata.txt", "r") as f: for line in f: tokens = self.tokenizer.encode(line.strip(), maxlength=self.blocksize) if len(tokens) < 2: continue examples.append(torch.tensor(tokens, dtype=torch.long)) return examples ```

Extending the Implementation

The implementation is designed to be modular and extensible:

1. Tokenizer Training: Add your own tokenizer training pipeline by extending the existing tokenizer classes
2. Inference: Use the model for inference by loading a pretrained checkpoint
3. Finetuning: Extend the finetuning pipeline to support Roberta-specific tasks

Model Checkpoints

The pretrained model checkpoints are saved in the output directory with the following structure:

out_dir/ iter-{step}.pt # Regular checkpoints best.pt # Best checkpoint based on validation loss

Notes

• The implementation uses Roberta's special tokens (mask token ID: 50264, pad token ID: 1)
• The default masking probability is 15% (configurable)
• The model supports both MLM pretraining and regular language modeling
• The implementation is compatible with the existing Lit-GPT training infrastructure

RoBERTa Finetuning

This document provides instructions for finetuning RoBERTa models on text classification tasks.

Installation

Make sure you have all the required dependencies installed:

bash pip install -r requirements.txt

Dataset Format

The finetuning script supports two types of datasets:

1. Toy Dataset (default): A simple binary classification dataset for movie reviews
2. AG News Dataset: A 4-class news classification dataset

Toy Dataset Format

The toy dataset should be organized as follows:

data/ test_data_finetune/ train.json test.json

Each JSON file should contain a list of examples, where each example has the following format:

json { "text": "The movie was fantastic!", "label": 1 }

AG News Dataset Format

The AG News dataset should be organized similarly, with each line in the JSON files containing a news article and its category label.

Finetuning

To finetune RoBERTa on the toy dataset:

bash python scripts/finetune_roberta.py \ --data_dir data/test_data_finetune \ --output_dir out/roberta_finetuned \ --num_labels 2 \ --batch_size 16 \ --learning_rate 2e-5 \ --num_epochs 3

To finetune on the AG News dataset:

bash python scripts/finetune_roberta.py \ --data_dir path/to/agnews \ --output_dir out/roberta_finetuned \ --num_labels 4 \ --batch_size 16 \ --learning_rate 2e-5 \ --num_epochs 3 \ --use_agnews

Command Line Arguments

• --data_dir: Directory containing the dataset (required)
• --output_dir: Output directory for the finetuned model (default: "out/roberta_finetuned")
• --model_name: Name of the pretrained RoBERTa model (default: "roberta-base")
• --num_labels: Number of classification labels (default: 2)
• --batch_size: Training batch size (default: 16)
• --learning_rate: Learning rate (default: 2e-5)
• --num_epochs: Number of training epochs (default: 3)
• --max_length: Maximum sequence length (default: 128)
• --use_agnews: Use AG News dataset instead of toy dataset (flag)

Evaluation

The script automatically evaluates the model on the test set after training. The following metrics are computed:

• Loss
• Accuracy
• F1 Score
• Precision
• Recall

The final metrics are printed to the console after training completes.

Model Output

The finetuned model is saved in the specified output directory with the following files:

• config.json: Model configuration
• pytorch_model.bin: Model weights
• tokenizer.json: Tokenizer configuration
• classifier.pt: Classification head weights

Example

Here's a complete example of finetuning on the toy dataset:

```bash

Create data directory

mkdir -p data/testdatafinetune

Copy the toy dataset files

cp data/testdatafinetune/train.json data/testdatafinetune/ cp data/testdatafinetune/test.json data/testdatafinetune/

Run finetuning

python scripts/finetuneroberta.py \ --datadir data/testdatafinetune \ --outputdir out/robertafinetuned \ --numlabels 2 \ --batchsize 16 \ --learningrate 2e-5 \ --numepochs 3 ```

The script will train the model and output evaluation metrics. The finetuned model will be saved in the out/roberta_finetuned directory.

⚡ LitGPT

20+ high-performance LLMs with recipes to pretrain, finetune, and deploy at scale.

✅ From scratch implementations      ✅ No abstractions         ✅ Beginner friendly
   ✅ Flash attention                   ✅ FSDP                    ✅ LoRA, QLoRA, Adapter
✅ Reduce GPU memory (fp4/8/16/32)   ✅ 1-1000+ GPUs/TPUs       ✅ 20+ LLMs

---

Quick start • Models • Finetune • Deploy • All workflows • Features • Recipes (YAML) • Lightning AI • Tutorials

 

 

Use, finetune, pretrain, and deploy LLMs Lightning fast ⚡⚡

Every LLM is implemented from scratch with no abstractions and full control, making them blazing fast, minimal, and performant at enterprise scale.

✅ Enterprise ready - Apache 2.0 for unlimited enterprise use.
✅ Developer friendly - Easy debugging with no abstraction layers and single file implementations.
✅ Optimized performance - Models designed to maximize performance, reduce costs, and speed up training.
✅ Proven recipes - Highly-optimized training/finetuning recipes tested at enterprise scale.

 

Quick start

Install LitGPT pip install 'litgpt[extra]'

Load and use any of the 20+ LLMs: ```python from litgpt import LLM

llm = LLM.load("microsoft/phi-2") text = llm.generate("Fix the spelling: Every fall, the family goes to the mountains.") print(text)

Corrected Sentence: Every fall, the family goes to the mountains.

```

 

✅ Optimized for fast inference
✅ Quantization
✅ Runs on low-memory GPUs
✅ No layers of internal abstractions
✅ Optimized for production scale

Advanced install options

Install from source: ```bash git clone https://github.com/Lightning-AI/litgpt cd litgpt pip install -e '.[all]' ```

Explore the full Python API docs.

 

---

Choose from 20+ LLMs

Every model is written from scratch to maximize performance and remove layers of abstraction:

| Model | Model size | Author | Reference | |----|----|----|----| | Llama 3, 3.1, 3.2, 3.3 | 1B, 3B, 8B, 70B, 405B | Meta AI | Meta AI 2024 | | Code Llama | 7B, 13B, 34B, 70B | Meta AI | Rozière et al. 2023 | | CodeGemma | 7B | Google | Google Team, Google Deepmind | | Gemma 2 | 2B, 9B, 27B | Google | Google Team, Google Deepmind | | Phi 4 | 14B | Microsoft Research | Abdin et al. 2024 | | Qwen2.5 | 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B | Alibaba Group | Qwen Team 2024 | | Qwen2.5 Coder | 0.5B, 1.5B, 3B, 7B, 14B, 32B | Alibaba Group | Hui, Binyuan et al. 2024 | | R1 Distill Llama | 8B, 70B | DeepSeek AI | DeepSeek AI 2025 | | ... | ... | ... | ... |

See full list of 20+ LLMs

  #### All models | Model | Model size | Author | Reference | |----|----|----|----| | CodeGemma | 7B | Google | [Google Team, Google Deepmind](https://ai.google.dev/gemma/docs/codegemma) | | Code Llama | 7B, 13B, 34B, 70B | Meta AI | [Rozière et al. 2023](https://arxiv.org/abs/2308.12950) | | Falcon | 7B, 40B, 180B | TII UAE | [TII 2023](https://falconllm.tii.ae) | | Falcon 3 | 1B, 3B, 7B, 10B | TII UAE | [TII 2024](https://huggingface.co/blog/falcon3) | | FreeWilly2 (Stable Beluga 2) | 70B | Stability AI | [Stability AI 2023](https://stability.ai/blog/stable-beluga-large-instruction-fine-tuned-models) | | Function Calling Llama 2 | 7B | Trelis | [Trelis et al. 2023](https://huggingface.co/Trelis/Llama-2-7b-chat-hf-function-calling-v2) | | Gemma | 2B, 7B | Google | [Google Team, Google Deepmind](https://storage.googleapis.com/deepmind-media/gemma/gemma-report.pdf) | | Gemma 2 | 9B, 27B | Google | [Google Team, Google Deepmind](https://storage.googleapis.com/deepmind-media/gemma/gemma-2-report.pdf) | | Gemma 3 | 1B, 4B, 12B, 27B | Google | [Google Team, Google Deepmind](https://arxiv.org/pdf/2503.19786) | | Llama 2 | 7B, 13B, 70B | Meta AI | [Touvron et al. 2023](https://arxiv.org/abs/2307.09288) | | Llama 3.1 | 8B, 70B | Meta AI | [Meta AI 2024](https://github.com/meta-llama/llama3) | | Llama 3.2 | 1B, 3B | Meta AI | [Meta AI 2024](https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/) | | Llama 3.3 | 70B | Meta AI | [Meta AI 2024](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) | | Mathstral | 7B | Mistral AI | [Mistral AI 2024](https://mistral.ai/news/mathstral/) | | MicroLlama | 300M | Ken Wang | [MicroLlama repo](https://github.com/keeeeenw/MicroLlama) | | Mixtral MoE | 8x7B | Mistral AI | [Mistral AI 2023](https://mistral.ai/news/mixtral-of-experts/) | | Mistral | 7B, 123B | Mistral AI | [Mistral AI 2023](https://mistral.ai/news/announcing-mistral-7b/) | | Mixtral MoE | 8x22B | Mistral AI | [Mistral AI 2024](https://mistral.ai/news/mixtral-8x22b/) | | OLMo | 1B, 7B | Allen Institute for AI (AI2) | [Groeneveld et al. 2024](https://aclanthology.org/2024.acl-long.841/) | | OpenLLaMA | 3B, 7B, 13B | OpenLM Research | [Geng & Liu 2023](https://github.com/openlm-research/open_llama) | | Phi 1.5 & 2 | 1.3B, 2.7B | Microsoft Research | [Li et al. 2023](https://arxiv.org/abs/2309.05463) | | Phi 3 | 3.8B | Microsoft Research | [Abdin et al. 2024](https://arxiv.org/abs/2404.14219) | | Phi 4 | 14B | Microsoft Research | [Abdin et al. 2024](https://arxiv.org/abs/2412.08905) | | Phi 4 Mini Instruct | 3.8B | Microsoft Research | [Microsoft 2025](https://arxiv.org/abs/2503.01743) | | Platypus | 7B, 13B, 70B | Lee et al. | [Lee, Hunter, and Ruiz 2023](https://arxiv.org/abs/2308.07317) | | Pythia | {14,31,70,160,410}M, {1,1.4,2.8,6.9,12}B | EleutherAI | [Biderman et al. 2023](https://arxiv.org/abs/2304.01373) | | Qwen2.5 | 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B | Alibaba Group | [Qwen Team 2024](https://qwenlm.github.io/blog/qwen2.5/) | | Qwen2.5 Coder | 0.5B, 1.5B, 3B, 7B, 14B, 32B | Alibaba Group | [Hui, Binyuan et al. 2024](https://arxiv.org/abs/2409.12186) | | Qwen2.5 Math | 1.5B, 7B, 72B | Alibaba Group | [An, Yang et al. 2024](https://arxiv.org/abs/2409.12122) | | QwQ | 32B | Alibaba Group | [Qwen Team 2025](https://qwenlm.github.io/blog/qwq-32b/) | | QwQ-Preview | 32B | Alibaba Group | [Qwen Team 2024](https://qwenlm.github.io/blog/qwq-32b-preview/) | | Roberta-base | 125M | Facebook AI | [Liu et al. 2019](https://arxiv.org/abs/1907.11692) | | R1 Distill Llama | 8B, 70B | DeepSeek AI | [DeepSeek AI 2025](https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf) | | SmolLM2 | 135M, 360M, 1.7B | Hugging Face | [Hugging Face 2024](https://github.com/huggingface/smollm) | | Salamandra | 2B, 7B | Barcelona Supercomputing Centre | [BSC-LTC 2024](https://github.com/BSC-LTC/salamandra) | | StableCode | 3B | Stability AI | [Stability AI 2023](https://stability.ai/blog/stablecode-llm-generative-ai-coding) | | StableLM | 3B, 7B | Stability AI | [Stability AI 2023](https://github.com/Stability-AI/StableLM) | | StableLM Zephyr | 3B | Stability AI | [Stability AI 2023](https://stability.ai/blog/stablecode-llm-generative-ai-coding) | | TinyLlama | 1.1B | Zhang et al. | [Zhang et al. 2023](https://github.com/jzhang38/TinyLlama) | **Tip**: You can list all available models by running the `litgpt download list` command.

 

---

Workflows

Finetune • Pretrain • Continued pretraining • Evaluate • Deploy • Test

 

Use the command line interface to run advanced workflows such as pretraining or finetuning on your own data.

All workflows

After installing LitGPT, select the model and workflow to run (finetune, pretrain, evaluate, deploy, etc...):

```bash

litgpt [action] [model]

litgpt serve meta-llama/Llama-3.2-3B-Instruct litgpt finetune meta-llama/Llama-3.2-3B-Instruct litgpt pretrain meta-llama/Llama-3.2-3B-Instruct litgpt chat meta-llama/Llama-3.2-3B-Instruct litgpt evaluate meta-llama/Llama-3.2-3B-Instruct ```

 

---

Finetune an LLM

 

Finetuning is the process of taking a pretrained AI model and further training it on a smaller, specialized dataset tailored to a specific task or application.

 

```bash

0) setup your dataset

curl -L https://huggingface.co/datasets/ksaw008/financealpaca/resolve/main/financealpaca.json -o mycustomdataset.json

1) Finetune a model (auto downloads weights)

litgpt finetune microsoft/phi-2 \ --data JSON \ --data.jsonpath mycustomdataset.json \ --data.valsplitfraction 0.1 \ --outdir out/custom-model

2) Test the model

litgpt chat out/custom-model/final

3) Deploy the model

litgpt serve out/custom-model/final ```

Read the full finetuning docs

 

---

Deploy an LLM

 

Deploy a pretrained or finetune LLM to use it in real-world applications. Deploy, automatically sets up a web server that can be accessed by a website or app.

```bash

deploy an out-of-the-box LLM

litgpt serve microsoft/phi-2

deploy your own trained model

litgpt serve path/to/microsoft/phi-2/checkpoint ```

Show code to query server:

  Test the server in a separate terminal and integrate the model API into your AI product: ```python # 3) Use the server (in a separate Python session) import requests, json response = requests.post( "http://127.0.0.1:8000/predict", json={"prompt": "Fix typos in the following sentence: Example input"} ) print(response.json()["output"]) ```

Read the full deploy docs.

 

---

Evaluate an LLM

Evaluate an LLM to test its performance on various tasks to see how well it understands and generates text. Simply put, we can evaluate things like how well would it do in college-level chemistry, coding, etc... (MMLU, Truthful QA, etc...)

bash litgpt evaluate microsoft/phi-2 --tasks 'truthfulqa_mc2,mmlu'

Read the full evaluation docs.

 

---

Test an LLM

 

Test how well the model works via an interactive chat. Use the chat command to chat, extract embeddings, etc...

Here's an example showing how to use the Phi-2 LLM: ```bash litgpt chat microsoft/phi-2

Prompt: What do Llamas eat? ```

Full code:

  ```bash # 1) List all supported LLMs litgpt download list # 2) Use a model (auto downloads weights) litgpt chat microsoft/phi-2 >> Prompt: What do Llamas eat? ``` The download of certain models requires an additional access token. You can read more about this in the [download](tutorials/download_model_weights.md#specific-models-and-access-tokens) documentation.

Read the full chat docs.

 

---

Pretrain an LLM

 

Pretraining is the process of training a language model from scratch on a large corpus of text data. LitGPT supports both causal language modeling (CLM) and masked language modeling (MLM) pretraining.

Causal Language Modeling (CLM)

```bash

Pretrain a model from scratch

litgpt pretrain meta-llama/Llama-3.2-3B-Instruct \ --data TinyLlama \ --out_dir out/pretrained-model ```

Masked Language Modeling (MLM)

For models like Roberta-base that use masked language modeling:

```bash

Pretrain a Roberta-base model with MLM

litgpt pretrain roberta-base \ --datadir path/to/testdata \ --outdir path/to/output \ --train.globalbatchsize 32 \ --train.microbatchsize 4 \ --train.maxtokens 1000000 \ --train.learningrate 1e-4 \ --train.lrwarmupsteps 2000 \ --train.saveinterval 1000 \ --eval.interval 1000 ```

The test dataset will be automatically created if it doesn't exist. You can also use your own dataset by creating a custom dataset class that inherits from TestMLMDataset.

Read the full pretraining docs

 

---

Continue pretraining an LLM

 

Continued pretraining is another way of finetuning that specializes an already pretrained model by training on custom data:

Show code:

  ```bash mkdir -p custom_texts curl https://www.gutenberg.org/cache/epub/24440/pg24440.txt --output custom_texts/book1.txt curl https://www.gutenberg.org/cache/epub/26393/pg26393.txt --output custom_texts/book2.txt # 1) Continue pretraining a model (auto downloads weights) litgpt pretrain EleutherAI/pythia-160m \ --tokenizer_dir EleutherAI/pythia-160m \ --initial_checkpoint_dir EleutherAI/pythia-160m \ --data TextFiles \ --data.train_data_path "custom_texts/" \ --train.max_tokens 10_000_000 \ --out_dir out/custom-model # 2) Test the model litgpt chat out/custom-model/final ```

Read the full continued pretraining docs

 

---

State-of-the-art features

✅ State-of-the-art optimizations: Flash Attention v2, multi-GPU support via fully-sharded data parallelism, optional CPU offloading, and TPU and XLA support.
✅ Pretrain, finetune, and deploy
✅ Reduce compute requirements with low-precision settings: FP16, BF16, and FP16/FP32 mixed.
✅ Lower memory requirements with quantization: 4-bit floats, 8-bit integers, and double quantization.
✅ Configuration files for great out-of-the-box performance.
✅ Parameter-efficient finetuning: LoRA, QLoRA, Adapter, and Adapter v2.
✅ Exporting to other popular model weight formats.
✅ Many popular datasets for pretraining and finetuning, and support for custom datasets.
✅ Readable and easy-to-modify code to experiment with the latest research ideas.

 

---

Training recipes

LitGPT comes with validated recipes (YAML configs) to train models under different conditions. We've generated these recipes based on the parameters we found to perform the best for different training conditions.

Browse all training recipes here.

Example

bash litgpt finetune \ --config https://raw.githubusercontent.com/Lightning-AI/litgpt/main/config_hub/finetune/llama-2-7b/lora.yaml

✅ Use configs to customize training

Configs let you customize training for all granular parameters like:

```yaml

The path to the base model's checkpoint directory to load for finetuning. (type: , default: checkpoints/stabilityai/stablelm-base-alpha-3b)

checkpoint_dir: checkpoints/meta-llama/Llama-2-7b-hf

Directory in which to save checkpoints and logs. (type: , default: out/lora)

out_dir: out/finetune/qlora-llama2-7b

The precision to use for finetuning. Possible choices: "bf16-true", "bf16-mixed", "32-true". (type: Optional[str], default: null)

precision: bf16-true

... ```

✅ Example: LoRA finetuning config

  ```yaml # The path to the base model's checkpoint directory to load for finetuning. (type: , default: checkpoints/stabilityai/stablelm-base-alpha-3b) checkpoint_dir: checkpoints/meta-llama/Llama-2-7b-hf # Directory in which to save checkpoints and logs. (type: , default: out/lora) out_dir: out/finetune/qlora-llama2-7b # The precision to use for finetuning. Possible choices: "bf16-true", "bf16-mixed", "32-true". (type: Optional[str], default: null) precision: bf16-true # If set, quantize the model with this algorithm. See ``tutorials/quantize.md`` for more information. (type: Optional[Literal['nf4', 'nf4-dq', 'fp4', 'fp4-dq', 'int8-training']], default: null) quantize: bnb.nf4 # How many devices/GPUs to use. (type: Union[int, str], default: 1) devices: 1 # How many nodes to use. (type: int, default: 1) num_nodes: 1 # The LoRA rank. (type: int, default: 8) lora_r: 32 # The LoRA alpha. (type: int, default: 16) lora_alpha: 16 # The LoRA dropout value. (type: float, default: 0.05) lora_dropout: 0.05 # Whether to apply LoRA to the query weights in attention. (type: bool, default: True) lora_query: true # Whether to apply LoRA to the key weights in attention. (type: bool, default: False) lora_key: false # Whether to apply LoRA to the value weights in attention. (type: bool, default: True) lora_value: true # Whether to apply LoRA to the output projection in the attention block. (type: bool, default: False) lora_projection: false # Whether to apply LoRA to the weights of the MLP in the attention block. (type: bool, default: False) lora_mlp: false # Whether to apply LoRA to output head in GPT. (type: bool, default: False) lora_head: false # Data-related arguments. If not provided, the default is ``litgpt.data.Alpaca``. data: class_path: litgpt.data.Alpaca2k init_args: mask_prompt: false val_split_fraction: 0.05 prompt_style: alpaca ignore_index: -100 seed: 42 num_workers: 4 download_dir: data/alpaca2k # Training-related arguments. See ``litgpt.args.TrainArgs`` for details train: # Number of optimizer steps between saving checkpoints (type: Optional[int], default: 1000) save_interval: 200 # Number of iterations between logging calls (type: int, default: 1) log_interval: 1 # Number of samples between optimizer steps across data-parallel ranks (type: int, default: 128) global_batch_size: 8 # Number of samples per data-parallel rank (type: int, default: 4) micro_batch_size: 2 # Number of iterations with learning rate warmup active (type: int, default: 100) lr_warmup_steps: 10 # Number of epochs to train on (type: Optional[int], default: 5) epochs: 4 # Total number of tokens to train on (type: Optional[int], default: null) max_tokens: # Limits the number of optimizer steps to run (type: Optional[int], default: null) max_steps: # Limits the length of samples (type: Optional[int], default: null) max_seq_length: 512 # Whether to tie the embedding weights with the language modeling head weights (type: Optional[bool], default: null) tie_embeddings: # (type: float, default: 0.0003) learning_rate: 0.0002 # (type: float, default: 0.02) weight_decay: 0.0 # (type: float, default: 0.9) beta1: 0.9 # (type: float, default: 0.95) beta2: 0.95 # (type: Optional[float], default: null) max_norm: # (type: float, default: 6e-05) min_lr: 6.0e-05 # Evaluation-related arguments. See ``litgpt.args.EvalArgs`` for details eval: # Number of optimizer steps between evaluation calls (type: int, default: 100) interval: 100 # Number of tokens to generate (type: Optional[int], default: 100) max_new_tokens: 100 # Number of iterations (type: int, default: 100) max_iters: 100 # The name of the logger to send metrics to. (type: Literal['wandb', 'tensorboard', 'csv'], default: csv) logger_name: csv # The random seed to use for reproducibility. (type: int, default: 1337) seed: 1337 ```

✅ Override any parameter in the CLI:

```bash litgpt finetune \ --config https://raw.githubusercontent.com/Lightning-AI/litgpt/main/config_hub/finetune/llama-2-7b/lora.yaml \ --lora_r 4 ```

 

---

Project highlights

LitGPT powers many great AI projects, initiatives, challenges and of course enterprises. Please submit a pull request to be considered for a feature.

📊 SAMBA: Simple Hybrid State Space Models for Efficient Unlimited Context Language Modeling

The [Samba](https://github.com/microsoft/Samba) project by researchers at Microsoft is built on top of the LitGPT code base and combines state space models with sliding window attention, which outperforms pure state space models.

🏆 NeurIPS 2023 Large Language Model Efficiency Challenge: 1 LLM + 1 GPU + 1 Day

The LitGPT repository was the official starter kit for the [NeurIPS 2023 LLM Efficiency Challenge](https://llm-efficiency-challenge.github.io), which is a competition focused on finetuning an existing non-instruction tuned LLM for 24 hours on a single GPU.

🦙 TinyLlama: An Open-Source Small Language Model

LitGPT powered the [TinyLlama project](https://github.com/jzhang38/TinyLlama) and [TinyLlama: An Open-Source Small Language Model](https://arxiv.org/abs/2401.02385) research paper.

🍪 MicroLlama: MicroLlama-300M

[MicroLlama](https://github.com/keeeeenw/MicroLlama) is a 300M Llama model pretrained on 50B tokens powered by TinyLlama and LitGPT.

🔬 Pre-training Small Base LMs with Fewer Tokens

The research paper ["Pre-training Small Base LMs with Fewer Tokens"](https://arxiv.org/abs/2404.08634), which utilizes LitGPT, develops smaller base language models by inheriting a few transformer blocks from larger models and training on a tiny fraction of the data used by the larger models. It demonstrates that these smaller models can perform comparably to larger models despite using significantly less training data and resources.

 

---

Community

We welcome all individual contributors, regardless of their level of experience or hardware. Your contributions are valuable, and we are excited to see what you can accomplish in this collaborative and supportive environment.

• Request a feature
• Submit your first contribution
• Join our Discord

 

Tutorials

🚀 Get started
⚡️ Finetuning, incl. LoRA, QLoRA, and Adapters
🤖 Pretraining
💬 Model evaluation
📘 Supported and custom datasets
🧹 Quantization
🤯 Tips for dealing with out-of-memory (OOM) errors
🧑🏽‍💻 Using cloud TPUs

 

---

Acknowledgments

This implementation extends on Lit-LLaMA and nanoGPT, and it's powered by Lightning Fabric ⚡.

• @karpathy for nanoGPT
• @EleutherAI for GPT-NeoX and the Evaluation Harness
• @TimDettmers for bitsandbytes
• @Microsoft for LoRA
• @tridao for Flash Attention 2

License

LitGPT is released under the Apache 2.0 license.

Citation

If you use LitGPT in your research, please cite the following work:

bibtex @misc{litgpt-2023, author = {Lightning AI}, title = {LitGPT}, howpublished = {\url{https://github.com/Lightning-AI/litgpt}}, year = {2023}, }