|
|
--- |
|
|
base_model: google/gemma-3-270m-it |
|
|
library_name: peft |
|
|
pipeline_tag: text-generation |
|
|
tags: |
|
|
- base_model:adapter:google/gemma-3-270m-it |
|
|
- lora |
|
|
- transformers |
|
|
--- |
|
|
|
|
|
🧠 MODEL CARD — DogeAI-v1.0-instruct |
|
|
Model Details |
|
|
|
|
|
Model Description |
|
|
DogeAI-v1.0-instruct is an early-stage instruction-following language model fine-tuned for conversational use and experimentation. This version is intended as a proof of concept (v1) and focuses on language generation rather than reliable logical reasoning. |
|
|
|
|
|
Developed by: Arthur(loboGOAT) |
|
|
Funded by: Independent / Community-driven |
|
|
Shared by: Arthur(loboGOAT) |
|
|
Model type: Small Instruction-Tuned Language Model |
|
|
Language(s): Portuguese (primary), multilingual tendencies inherited from base model |
|
|
License: Apache 2.0 (or the same license as the base model, if different) |
|
|
Finetuned from model: Gemma-3-270M-it |
|
|
|
|
|
Model Sources |
|
|
|
|
|
Repository: loboGOAT/DogeAI-v1.0-instruct |
|
|
Paper: Not available |
|
|
Demo: Not available |
|
|
|
|
|
Uses |
|
|
Direct Use |
|
|
|
|
|
Conversational experiments |
|
|
|
|
|
Text generation and rewriting |
|
|
|
|
|
Prompt testing and evaluation |
|
|
|
|
|
Educational use to study limitations of small LLMs |
|
|
|
|
|
Downstream Use (Optional) |
|
|
|
|
|
Further fine-tuning |
|
|
|
|
|
Research on alignment, reasoning, and instruction-following |
|
|
|
|
|
Benchmarking small models |
|
|
|
|
|
Out-of-Scope Use |
|
|
|
|
|
Tasks requiring reliable logical reasoning |
|
|
|
|
|
Mathematical proof or formal logic |
|
|
|
|
|
Decision-making systems |
|
|
|
|
|
Safety-critical or automated validation tasks |
|
|
|
|
|
Recommendations |
|
|
|
|
|
This model should not be relied upon for reasoning-intensive tasks. |
|
|
Users are encouraged to treat DogeAI-v1.0-instruct as an experimental model and expect occasional logical inconsistencies, multilingual drift, or overgeneration. |
|
|
|
|
|
Future versions aim to address these limitations through: |
|
|
|
|
|
cleaner datasets |
|
|
|
|
|
improved stopping criteria |
|
|
|
|
|
alternative base models |
|
|
|
|
|
How to Get Started with the Model |
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained("loboGOAT/DogeAI-v1.0-instruct") |
|
|
model = AutoModelForCausalLM.from_pretrained("loboGOAT/DogeAI-v1.0-instruct") |
|
|
|
|
|
inputs = tokenizer("Olá! Vamos conversar?", return_tensors="pt") |
|
|
outputs = model.generate( |
|
|
**inputs, |
|
|
max_new_tokens=128, |
|
|
temperature=0.65, |
|
|
top_p=0.95 |
|
|
) |
|
|
|
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
|
|
|
|
|
Training Details |
|
|
Training Data |
|
|
|
|
|
The model was fine-tuned on a custom instruction-style dataset, primarily in Portuguese, designed to encourage conversational responses. |
|
|
The dataset does not focus on formal logic or structured reasoning. |
|
|
|
|
|
Training Procedure |
|
|
|
|
|
Preprocessing |
|
|
|
|
|
Instruction–response formatting |
|
|
|
|
|
Text normalization |
|
|
|
|
|
No explicit chain-of-thought supervision |
|
|
|
|
|
Training Hyperparameters |
|
|
|
|
|
Training regime: Supervised fine-tuning (SFT) |
|
|
PEFT: Yes (LoRA-based fine-tuning) |
|
|
|
|
|
Evaluation |
|
|
Testing Data |
|
|
|
|
|
Manual testing and prompt-based evaluation. |
|
|
|
|
|
Factors |
|
|
|
|
|
Logical consistency |
|
|
|
|
|
Instruction-following |
|
|
|
|
|
Language fluency |
|
|
|
|
|
Metrics |
|
|
|
|
|
No automated benchmarks were used for this version. |
|
|
|
|
|
Results |
|
|
|
|
|
Strong conversational fluency for model size |
|
|
|
|
|
Inconsistent logical reasoning |
|
|
|
|
|
Occasional overgeneration beyond intended response |
|
|
|
|
|
Summary |
|
|
Model Examination |
|
|
|
|
|
DogeAI-v1.0-instruct demonstrates the strengths and limitations of small instruction-tuned language models. |
|
|
While capable of natural conversation, it lacks robust reasoning abilities, which will be a focus of future iterations. |
|
|
|
|
|
Environmental Impact |
|
|
|
|
|
Hardware Type: Consumer GPU / Local Machine |
|
|
Hours used: Low |
|
|
Cloud Provider: None |
|
|
Compute Region: Local |
|
|
Carbon Emitted: Negligible |
|
|
|
|
|
Technical Specifications |
|
|
Model Architecture and Objective |
|
|
|
|
|
Decoder-only Transformer |
|
|
|
|
|
Next-token prediction |
|
|
|
|
|
Instruction-following objective |
|
|
|
|
|
Compute Infrastructure |
|
|
|
|
|
Local training environment. |
|
|
|
|
|
Hardware |
|
|
|
|
|
Consumer-grade GPU / CPU |
|
|
|
|
|
Software |
|
|
|
|
|
Transformers |
|
|
|
|
|
PEFT 0.18.0 |
|
|
|
|
|
PyTorch |
|
|
|
|
|
Citation |
|
|
|
|
|
BibTeX: |
|
|
|
|
|
@misc{dogeai_v1_2025, |
|
|
title={DogeAI-v1.0-instruct}, |
|
|
author={Arthur}, |
|
|
year={2025}, |
|
|
note={Early experimental instruction-tuned language model} |
|
|
} |
|
|
|
|
|
|
|
|
APA: |
|
|
Arthur (2025). DogeAI-v1.0-instruct: An experimental instruction-tuned language model. |
|
|
|
|
|
Model Card Authors |
|
|
|
|
|
Arthur |
|
|
|
|
|
Model Card Contact |
|
|
|
|
|
(your Hugging Face profile or GitHub) |
