🧠 dpp-gpt v2.1 Base (260M)

(🇺🇸 English / 🇷🇺 Русский)

⚠️ Note: This is a base foundation model. It has only undergone pre-training and has not been fine-tuned for general chat. For the instruction-following version, please download dpp-gpt-v2.1-pro-260m.

This is a compact foundation language model with 259M parameters, trained entirely from scratch. Despite being a raw base model, it possesses strong native mathematical, tokenization, and translation capabilities embedded directly into its pre-training weights.

⚙️ Model Details

  • Parameters: 259M
  • Layers / Hidden Size / Heads: 20 / 1024 / 16
  • Context Length: 4096 tokens
  • Vocabulary Size: 16,384
  • Type: Base (Pre-trained foundation model)
  • Format: GGUF / PyTorch (.pth)
  • License: Apache 2.0

📊 Pre-training Data

  • Dataset Size: 11.8 Billion tokens (~45.5 tokens/parameter, batch size 512k).
  • Languages: Balanced multi-lingual corpus consisting of Russian, English, and French data.

💡 Prompt Guide & Real Examples

As a raw base model, it is highly sensitive to syntax. Mathematical and textual formatting was baked into the pre-training using ChatML syntax. Use strictly the following structures to get accurate results:

🌍 Basic Translation

⚠️ Tip: This model is highly sensitive to the user's prompt formatting. To get a stable and correct translation, you must start your prompt with a capital letter and experiment with punctuation (sometimes adding or removing a period at the end changes the output). Also, cross-lingual prompts (e.g., asking in Russian to translate an English text into French) might yield less natural results.

<|im_start|>user
Переведи на английский: Я тебя люблю.<|im_end|>
<|im_start|>assistant
I love you.

🧮 Math Calculation (Chain-of-Thought)

To activate step-by-step math solving, use the [THINK] token:

<|im_start|>user
[THINK] 1523 - 659 + 234<|im_end|>
<|im_start|>assistant

(The model will continue by breaking down the calculation into Thousands, Hundreds, Tens, and Units).

🔢 Counting Total Letters in a Word

<|im_start|>user
How many letters in "apple"?<|im_end|>
<|im_start|>assistant
Breaking down the word "apple" into letters:
a-p-p-l-e
a - 1
p - 2
p - 3
l - 4
e - 5
The word apple has 5 letters.

🎯 Counting Specific Letters

<|im_start|>user
How many letters "r" in "strawberry"?<|im_end|>
<|im_start|>assistant
Breaking down the word "strawberry" into letters:
s-t-r-a-w-b-e-r-r-y
s - 0
t - 0
r - 1
a - 1
w - 1
b - 1
e - 1
r - 2
r - 3
y - 3

Answer: 3

🔤 Spelling a Word Letter-by-Letter

⚠️ Important: When asking the model to simply spell a word, you MUST put a period . at the end of your prompt, otherwise the logic will break.

<|im_start|>user
Spell the word apple.<|im_end|>
<|im_start|>assistant
a-p-p-l-e

🇷🇺 Описание на русском

⚠️ Внимание: Это базовая (foundation) модель. Она прошла только этап pre-training и не обучалась свободному ведению диалога. Если вам нужна готовая чат-версия, скачайте dpp-gpt-v2.1-pro-260m.

Это базовая компактная языковая модель на 259М параметров, обученная полностью с нуля. Примечательно, что на этапе предобучения модель продемонстрировала мощные навыки счета, работы с текстом и базового перевода (в этой версии базовый математический аппарат работает даже точнее и стабильнее, чем в последующих SFT-модификациях).

⚙️ Детали модели

  • Параметры: 259M
  • Слои / Размерность / Головы: 20 / 1024 / 16
  • Контекст: 4096 токенов
  • Словарь: 16,384 токена
  • Формат: GGUF / PyTorch (.pth)
  • Лицензия: Apache 2.0

💡 Важное руководство по промптам (Prompt Guide)

Так как это сырая базовая модель, она крайне чувствительна к синтаксису запроса. Если формат нарушен — логика генерации сломается. Разметка зашивалась в претрейн через формат ChatML, поэтому используйте строго следующие конструкции:

🌍 Базовый перевод

⚠️ Совет: Модель крайне чувствительна к формату вашего запроса. Чтобы получить стабильный и правильный перевод, пользователю необходимо начинать фразу с заглавной буквы, а также иногда экспериментировать с пунктуацией (наличие или отсутствие точки в конце запроса может полностью поменять результат). Also, ей тяжелее даются кросс-языковые задачи (например, просьба на русском перевести английский текст на французский).

<|im_start|>user
Translate to russian: I love you.<|im_end|>
<|im_start|>assistant
Я люблю тебя.

🧮 Математический расчет (Chain-of-Thought)

Для активации пошагового решения используйте тег [THINK]:

<|im_start|>user
[THINK] 1523 - 659 + 234<|im_end|>
<|im_start|>assistant

(Модель сама продолжит текст, расписав вычитание и сложение по разрядам: Milliers, Centaines, Dizaines, Unités).

🔢 Подсчет всех букв в слове

<|im_start|>user
Сколько букв в слове яблоко?<|im_end|>
<|im_start|>assistant
Разбивка слова "яблоко" на буквы:
я-б-л-о-к-о
я - 1
б - 2
л - 3
о - 4
к - 5
о - 6
В слове яблоко 6 букв.

🎯 Подсчет конкретной буквы

<|im_start|>user
Сколько букв "а" в слове абстракция?<|im_end|>
<|im_start|>assistant
Разбивка слова "абстракция" на буквы:
а-б-с-т-р-а-к-ц-и-я
а - 1
б - 1
с - 1
т - 1
р - 1
а - 2
к - 2
ц - 2
и - 2
я - 2

Ответ: 2

🔤 Побуквенный разбор слов

⚠️ Важно: При запросе на обычную разбивку по буквам, в конце промпта обязательно должна стоять точка .. Без неё модель может выдать некорректный результат.

<|im_start|>user
напиши слово яблоко по буквам.<|im_end|>
<|im_start|>assistant
я-б-л-о-к-о

Hotfix

The GGUF files in this repository were re-uploaded as a hotfix.

An LM Studio update changed the way special tokens are handled, which made the previously published GGUF files generate broken output. The conversion has been fixed and the quantizations here were rebuilt from the corrected model. The weights are unchanged — only the token metadata inside the GGUF files.

If you downloaded a GGUF from this repository before this commit, please download it again.

Downloads last month
53
GGUF
Model size
0.3B params
Architecture
llama
Hardware compatibility
Log In to add your hardware

2-bit

3-bit

4-bit

5-bit

6-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support