File size: 19,054 Bytes
6ef3642 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 |
---
base_model: nicholasKluge/TeenyTinyLlama-460m
co2_eq_emissions:
emissions: 41.1
geographical_location: Germany
hardware_used: NVIDIA A100-SXM4-40GB
source: CodeCarbon
training_type: pre-training
datasets:
- nicholasKluge/Pt-Corpus-Instruct
inference: true
language:
- pt
library_name: transformers
license: apache-2.0
metrics:
- perplexity
model-index:
- name: TeenyTinyLlama-460m
results:
- dataset:
args:
num_few_shot: 3
name: ENEM Challenge (No Images)
split: train
type: eduagarcia/enem_challenge
metrics:
- name: accuracy
type: acc
value: 20.15
source:
name: Open Portuguese LLM Leaderboard
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard?query=nicholasKluge/TeenyTinyLlama-460m
task:
name: Text Generation
type: text-generation
- dataset:
args:
num_few_shot: 3
name: BLUEX (No Images)
split: train
type: eduagarcia-temp/BLUEX_without_images
metrics:
- name: accuracy
type: acc
value: 25.73
source:
name: Open Portuguese LLM Leaderboard
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard?query=nicholasKluge/TeenyTinyLlama-460m
task:
name: Text Generation
type: text-generation
- dataset:
args:
num_few_shot: 3
name: OAB Exams
split: train
type: eduagarcia/oab_exams
metrics:
- name: accuracy
type: acc
value: 27.02
source:
name: Open Portuguese LLM Leaderboard
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard?query=nicholasKluge/TeenyTinyLlama-460m
task:
name: Text Generation
type: text-generation
- dataset:
args:
num_few_shot: 15
name: Assin2 RTE
split: test
type: assin2
metrics:
- name: f1-macro
type: f1_macro
value: 53.61
source:
name: Open Portuguese LLM Leaderboard
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard?query=nicholasKluge/TeenyTinyLlama-460m
task:
name: Text Generation
type: text-generation
- dataset:
args:
num_few_shot: 15
name: Assin2 STS
split: test
type: eduagarcia/portuguese_benchmark
metrics:
- name: pearson
type: pearson
value: 13.0
source:
name: Open Portuguese LLM Leaderboard
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard?query=nicholasKluge/TeenyTinyLlama-460m
task:
name: Text Generation
type: text-generation
- dataset:
args:
num_few_shot: 15
name: FaQuAD NLI
split: test
type: ruanchaves/faquad-nli
metrics:
- name: f1-macro
type: f1_macro
value: 46.41
source:
name: Open Portuguese LLM Leaderboard
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard?query=nicholasKluge/TeenyTinyLlama-460m
task:
name: Text Generation
type: text-generation
- dataset:
args:
num_few_shot: 25
name: HateBR Binary
split: test
type: ruanchaves/hatebr
metrics:
- name: f1-macro
type: f1_macro
value: 33.59
source:
name: Open Portuguese LLM Leaderboard
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard?query=nicholasKluge/TeenyTinyLlama-460m
task:
name: Text Generation
type: text-generation
- dataset:
args:
num_few_shot: 25
name: PT Hate Speech Binary
split: test
type: hate_speech_portuguese
metrics:
- name: f1-macro
type: f1_macro
value: 22.99
source:
name: Open Portuguese LLM Leaderboard
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard?query=nicholasKluge/TeenyTinyLlama-460m
task:
name: Text Generation
type: text-generation
- dataset:
args:
num_few_shot: 25
name: tweetSentBR
split: test
type: eduagarcia-temp/tweetsentbr
metrics:
- name: f1-macro
type: f1_macro
value: 17.28
source:
name: Open Portuguese LLM Leaderboard
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard?query=nicholasKluge/TeenyTinyLlama-460m
task:
name: Text Generation
type: text-generation
model_creator: nicholasKluge
model_name: TeenyTinyLlama-460m
pipeline_tag: text-generation
quantized_by: afrideva
tags:
- text-generation-inference
- gguf
- ggml
- quantized
widget:
- example_title: Exemplo
text: 'A PUCRS é uma universidade '
- example_title: Exemplo
text: A muitos anos atrás, em uma galáxia muito distante, vivia uma raça de
- example_title: Exemplo
text: Em meio a um escândalo, a frente parlamentar pediu ao Senador Silva para
---
# TeenyTinyLlama-460m-GGUF
Quantized GGUF model files for [TeenyTinyLlama-460m](https://huggingface.co/nicholasKluge/TeenyTinyLlama-460m) from [nicholasKluge](https://huggingface.co/nicholasKluge)
## Original Model Card:
# TeenyTinyLlama-460m
<img src="./logo.png" alt="A curious llama exploring a mushroom forest." height="200">
## Model Summary
Large language models (LLMs) have significantly advanced natural language processing, but their progress has yet to be equal across languages. While most LLMs are trained in high-resource languages like English, multilingual models generally underperform monolingual ones. Additionally, aspects of their multilingual foundation sometimes restrict the byproducts they produce, like computational demands and licensing regimes. Hence, we developed the _TeenyTinyLlama_ pair: two compact models for Brazilian Portuguese text generation.
Read our preprint on [ArXiv](https://arxiv.org/abs/2401.16640).
## Details
- **Architecture:** a Transformer-based model pre-trained via causal language modeling
- **Size:** 468,239,360 parameters
- **Context length:** 2048 tokens
- **Dataset:** [Pt-Corpus Instruct](https://huggingface.co/datasets/nicholasKluge/Pt-Corpus-Instruct) (6.2B tokens)
- **Language:** Portuguese
- **Number of steps:** 1,200,000
- **GPU:** 1 NVIDIA A100-SXM4-40GB
- **Training time**: ~ 280 hours
- **Emissions:** 41.1 KgCO2 (Germany)
- **Total energy consumption:** 115.69 kWh
This repository has the [source code](https://github.com/Nkluge-correa/TeenyTinyLlama) used to train this model. The main libraries used are:
- [Transformers](https://github.com/huggingface/transformers)
- [PyTorch](https://github.com/pytorch/pytorch)
- [Datasets](https://github.com/huggingface/datasets)
- [Tokenizers](https://github.com/huggingface/tokenizers)
- [Sentencepiece](https://github.com/google/sentencepiece)
- [Accelerate](https://github.com/huggingface/accelerate)
- [FlashAttention](https://github.com/Dao-AILab/flash-attention)
- [Codecarbon](https://github.com/mlco2/codecarbon)
## Intended Uses
The primary intended use of TeenyTinyLlama is to research the challenges related to developing language models for low-resource languages. Checkpoints saved during training are intended to provide a controlled setting for performing scientific experiments. You may also further fine-tune and adapt TeenyTinyLlama for deployment, as long as your use is following the Apache 2.0 license. If you decide to use pre-trained TeenyTinyLlama as a basis for your fine-tuned model, please conduct your own risk and bias assessment.
## Out-of-scope Use
TeenyTinyLlama is not intended for deployment. It is not a product and should not be used for human-facing interactions.
TeenyTinyLlama models are Brazilian Portuguese language only and are not suitable for translation or generating text in other languages.
TeenyTinyLlama has not been fine-tuned for downstream contexts in which language models are commonly deployed.
## Basic usage
Using the `pipeline`:
```python
from transformers import pipeline
generator = pipeline("text-generation", model="nicholasKluge/TeenyTinyLlama-460m")
completions = generator("Astronomia é a ciência", num_return_sequences=2, max_new_tokens=100)
for comp in completions:
print(f"🤖 {comp['generated_text']}")
```
Using the `AutoTokenizer` and `AutoModelForCausalLM`:
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
# Load model and the tokenizer
tokenizer = AutoTokenizer.from_pretrained("nicholasKluge/TeenyTinyLlama-460m", revision='main')
model = AutoModelForCausalLM.from_pretrained("nicholasKluge/TeenyTinyLlama-460m", revision='main')
# Pass the model to your device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.eval()
model.to(device)
# Tokenize the inputs and pass them to the device
inputs = tokenizer("Astronomia é a ciência", return_tensors="pt").to(device)
# Generate some text
completions = model.generate(**inputs, num_return_sequences=2, max_new_tokens=100)
# Print the generated text
for i, completion in enumerate(completions):
print(f'🤖 {tokenizer.decode(completion)}')
```
## Limitations
Like almost all other language models trained on large text datasets scraped from the web, the TTL pair exhibited behavior that does not make them an out-of-the-box solution to many real-world applications, especially those requiring factual, reliable, nontoxic text generation. Our models are all subject to the following:
- **Hallucinations:** This model can produce content that can be mistaken for truth but is, in fact, misleading or entirely false, i.e., hallucination.
- **Biases and Toxicity:** This model inherits the social and historical stereotypes from the data used to train it. Given these biases, the model can produce toxic content, i.e., harmful, offensive, or detrimental to individuals, groups, or communities.
- **Unreliable Code:** The model may produce incorrect code snippets and statements. These code generations should not be treated as suggestions or accurate solutions.
- **Language Limitations:** The model is primarily designed to understand standard Brazilian Portuguese. Other languages might challenge its comprehension, leading to potential misinterpretations or errors in response.
- **Repetition and Verbosity:** The model may get stuck on repetition loops (especially if the repetition penalty during generations is set to a meager value) or produce verbose responses unrelated to the prompt it was given.
Hence, even though our models are released with a permissive license, we urge users to perform their risk analysis on these models if intending to use them for real-world applications and also have humans moderating the outputs of these models in applications where they will interact with an audience, guaranteeing users are always aware they are interacting with a language model.
## Evaluations
During our training runs, both models showed consistent convergence. At no point did our evaluation curves show signs of overfitting or saturation. In the case of our 460m parameter model, we intentionally trained past the optimal point by approximately 75,000 steps to assess if there were any signs of saturation, but our evaluations consistently gave better results. We hypothesize that our models are under-trained but can improve if further trained to pass the Chinchilla optimal range.
| Processed Tokens | Perplexity | Energy Consumption (kWh) | Emissions (KgCO2eq) |
|------------------|------------|---------------------------|----------------------|
| 8.1M | 20.49 | 9.40 | 3.34 |
| 1.6B | 16.90 | 18.82 | 6.70 |
| 2.4B | 15.43 | 28.59 | 10.16 |
| 3.2B | 14.64 | 38.20 | 13.57 |
| 4.0B | 14.08 | 48.04 | 17.07 |
| 4.9B | 13.61 | 57.74 | 20.52 |
| 5.7B | 13.25 | 67.32 | 23.92 |
| 6.5B | 12.87 | 76.84 | 27.30 |
| 7.3B | 12.57 | 86.40 | 30.70 |
| 8.1B | 12.27 | 96.19 | 34.18 |
| 9.0B | 11.96 | 106.06 | 37.70 |
| 9.8B | 11.77 | 115.69 | 41.31 |
## Benchmarks
Evaluations on benchmarks were performed using the [Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) (by [EleutherAI](https://www.eleuther.ai/)). [Laiviet](https://github.com/laiviet/lm-evaluation-harness) translated the tasks from the LM-Evaluation-Harness we used. The results of models marked with an "*" were extracted from the [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard).
| | **ARC** | **HellaSwag** | **MMLU** | **TruthfulQA** | **Average** |
|------------------|-----------|---------------|-----------|----------------|-------------|
| Pythia-410m | 24.83* | 41.29* | 25.99* | 40.95* | 33.26 |
| **TTL-460m** | 29.40 | 33.00 | 28.55 | 41.10 | 33.01 |
| Bloom-560m | 24.74* | 37.15* | 24.22* | 42.44* | 32.13 |
| Xglm-564M | 25.56 | 34.64* | 25.18* | 42.53 | 31.97 |
| OPT-350m | 23.55* | 36.73* | 26.02* | 40.83* | 31.78 |
| **TTL-160m** | 26.15 | 29.29 | 28.11 | 41.12 | 31.16 |
| Pythia-160m | 24.06* | 31.39* | 24.86* | 44.34* | 31.16 |
| OPT-125m | 22.87* | 31.47* | 26.02* | 42.87* | 30.80 |
| GPorTuguese-2 | 22.48 | 29.62 | 27.36 | 41.44 | 30.22 |
| Gpt2-small | 21.48* | 31.60* | 25.79* | 40.65* | 29.97 |
| Multilingual GPT | 23.81 | 26.37* | 25.17* | 39.62 | 28.73 |
Evaluations on Brazilian Portuguese benchmarks were performed using a [Portuguese implementation of the EleutherAI LM Evaluation Harness](https://github.com/eduagarcia/lm-evaluation-harness-pt) (created by [Eduardo Garcia](https://github.com/eduagarcia/lm-evaluation-harness-pt)).
| | **ASSIN2 RTE** | **ASSIN2 STS** | **BLUEX** | **ENEM** | **FAQUAD NLI** | **HateBR** | **OAB Exams** | **Average** |
|----------------|----------------|----------------|-----------|----------|----------------|------------|---------------|-------------|
| Qwen-1.8B | 64.83 | 19.53 | 26.15 | 30.23 | 43.97 | 33.33 | 27.20 | 35.03 |
| TinyLlama-1.1B | 58.93 | 13.57 | 22.81 | 22.25 | 43.97 | 36.92 | 23.64 | 31.72 |
| **TTL-460m** | 53.93 | 12.66 | 22.81 | 19.87 | 49.01 | 33.59 | 27.06 | 31.27 |
| XGLM-564m | 49.61 | 22.91 | 19.61 | 19.38 | 43.97 | 33.99 | 23.42 | 30.41 |
| Bloom-1b7 | 53.60 | 4.81 | 21.42 | 18.96 | 43.97 | 34.89 | 23.05 | 28.67 |
| **TTL-160m** | 53.36 | 2.58 | 21.84 | 18.75 | 43.97 | 36.88 | 22.60 | 28.56 |
| OPT-125m | 39.77 | 2.00 | 21.84 | 17.42 | 43.97 | 47.04 | 22.78 | 27.83 |
| Pythia-160 | 33.33 | 12.81 | 16.13 | 16.66 | 50.36 | 41.09 | 22.82 | 27.60 |
| OLMo-1b | 34.12 | 9.28 | 18.92 | 20.29 | 43.97 | 41.33 | 22.96 | 27.26 |
| Bloom-560m | 33.33 | 8.48 | 18.92 | 19.03 | 43.97 | 37.07 | 23.05 | 26.26 |
| Pythia-410m | 33.33 | 4.80 | 19.47 | 19.45 | 43.97 | 33.33 | 23.01 | 25.33 |
| OPT-350m | 33.33 | 3.65 | 20.72 | 17.35 | 44.71 | 33.33 | 23.01 | 25.15 |
| GPT-2 small | 33.26 | 0.00 | 10.43 | 11.20 | 43.52 | 33.68 | 13.12 | 20.74 |
| GPorTuguese | 33.33 | 3.85 | 14.74 | 3.01 | 28.81 | 33.33 | 21.23 | 19.75 |
| Samba-1.1B | 33.33 | 1.30 | 8.07 | 10.22 | 17.72 | 35.79 | 15.03 | 17.35 |
## Fine-Tuning Comparisons
To further evaluate the downstream capabilities of our models, we decided to employ a basic fine-tuning procedure for our TTL pair on a subset of tasks from the Poeta benchmark. We apply the same procedure for comparison purposes on both [BERTimbau](https://huggingface.co/neuralmind/bert-base-portuguese-cased) models, given that they are also LLM trained from scratch in Brazilian Portuguese and have a similar size range to our models. We used these comparisons to assess if our pre-training runs produced LLM capable of producing good results ("good" here means "close to BERTimbau") when utilized for downstream applications.
| Models | IMDB | FaQuAD-NLI | HateBr | Assin2 | AgNews | Average |
|-----------------|-----------|------------|-----------|-----------|-----------|---------|
| BERTimbau-large | **93.58** | 92.26 | 91.57 | **88.97** | 94.11 | 92.10 |
| BERTimbau-small | 92.22 | **93.07** | 91.28 | 87.45 | 94.19 | 91.64 |
| **TTL-460m** | 91.64 | 91.18 | **92.28** | 86.43 | **94.42** | 91.19 |
| **TTL-160m** | 91.14 | 90.00 | 90.71 | 85.78 | 94.05 | 90.34 |
All the shown results are the higher accuracy scores achieved on the respective task test sets after fine-tuning the models on the training sets. All fine-tuning runs used the same hyperparameters, and the code implementation can be found in the [model cards](https://huggingface.co/nicholasKluge/TeenyTinyLlama-460m-HateBR) of our fine-tuned models.
## Cite as 🤗
```latex
@misc{correa24ttllama,
title = {TeenyTinyLlama: open-source tiny language models trained in Brazilian Portuguese},
author = {Corr{\^e}a, Nicholas Kluge and Falk, Sophia and Fatimah, Shiza and Sen, Aniket and De Oliveira, Nythamar},
journal={arXiv preprint arXiv:2401.16640},
year={2024}
}
```
## Funding
This repository was built as part of the RAIES ([Rede de Inteligência Artificial Ética e Segura](https://www.raies.org/)) initiative, a project supported by FAPERGS - ([Fundação de Amparo à Pesquisa do Estado do Rio Grande do Sul](https://fapergs.rs.gov.br/inicial)), Brazil.
## License
TeenyTinyLlama-460m is licensed under the Apache License, Version 2.0. See the [LICENSE](LICENSE) file for more details. |