MANTA-20M / README.md
average23's picture
docs: show word prediction output
a704e30 verified
|
Raw History Blame Contribute Delete
3.64 kB
metadata
license: apache-2.0
language:
  - ru
  - en
library_name: transformers
pipeline_tag: text-generation
tags:
  - MANTA
  - qwen2
  - autocomplete
  - word-completion
  - safetensors
model-index:
  - name: MANTA 20M
    results: []

MANTA 20M

v0.1.0 · Compact Russian conversational language model for next-word prediction and word completion.

Architecture

Parameter Value
Parameters 20,084,064
Decoder Qwen2-compatible, autoregressive
Transformer layers 8
Hidden size / intermediate size 352 / 1,024
Attention heads / KV heads 8 / 4
Vocabulary 24,000, ByteLevel BPE
Architecture context 1,024 tokens
Keyboard evaluation context 128 tokens
Weights FP32 Safetensors, tied embeddings

Training

The initial checkpoint was pretrained on Russian educational web text, general web text and Wikipedia. MANTA then received conversational adaptation: 779,333,270 target tokens in one corpus pass, followed by 41,074,231 target tokens over 2,239,218 windows. The final stage used 1,227,460 new Telegram messages and 1,009,803 replay windows from the earlier corpus.

Sources cover Telegram chats, conversational dialogue, forum text, slang and profanity, RU/EN code-switching, technical vocabulary, lexical normalization and a small news component. Final stage: AdamW, learning rate 3e-5, weight decay 0.01, batch size 16, 512-token packs, 5,212 updates, seed 42, FP32 parameters with CUDA BF16 autocast.

Evaluation

Model Parameters Next-word Top-1 Next-word Top-3 Completion Top-1 Completion Top-3
MANTA 20M 20,084,064 8.09% 15.74% 31.61% 52.59%
Keyboard LM 57M 57,321,984 5.96% 12.34% 26.68% 38.60%

Same Telegram cohort: 128 authors, 235 unique target positions, 940 next-word and 1,544 completion queries. Completion prefixes use one character and half the target word, deduplicating equal prefix lengths. Queries repeat across four history snapshots (0/10/100/500); MANTA uses the language model alone. Scoring uses NFC-normalized exact word match. Whole-word beam search: beam 10, 64 steps, 48-character word limit, 128 context tokens, 30-second ceiling, ten candidates before Top-3 scoring, FP32, TF32 disabled.

Usage

pip install https://huggingface.co/average23/MANTA-20M/resolve/v0.1.0/manta_keyboard-0.1.0-py3-none-any.whl
from manta import MantaKeyboard

keyboard = MantaKeyboard.from_pretrained("average23/MANTA-20M")
next_words = keyboard.next_word("ну это полный ")
completions = keyboard.complete("надо перезапустить ", prefix="док")
print([word["text"] for word in next_words])
print([word["text"] for word in completions])

Output:

['пиздец', 'бред', 'набор']
['докер', 'докерфайл', 'доку']

Both methods return up to three dictionaries with text (a complete word) and log_probability (explored word-boundary path mass). Completion preserves the typed prefix; completed context ends in whitespace. The default search has fixed beam and step budgets. Pass max_seconds=30 to use the benchmark deadline. Download once, then load the saved directory for offline inference.

Standard Transformers loading:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("average23/MANTA-20M", revision="v0.1.0", trust_remote_code=False)
tokenizer = AutoTokenizer.from_pretrained("average23/MANTA-20M", revision="v0.1.0", trust_remote_code=False)