Instructions to use TilQazyna/Til-1B-multilingual-base-GEC with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use TilQazyna/Til-1B-multilingual-base-GEC with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="TilQazyna/Til-1B-multilingual-base-GEC")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("TilQazyna/Til-1B-multilingual-base-GEC")
model = AutoModelForCausalLM.from_pretrained("TilQazyna/Til-1B-multilingual-base-GEC")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use TilQazyna/Til-1B-multilingual-base-GEC with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "TilQazyna/Til-1B-multilingual-base-GEC"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TilQazyna/Til-1B-multilingual-base-GEC",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/TilQazyna/Til-1B-multilingual-base-GEC

SGLang

How to use TilQazyna/Til-1B-multilingual-base-GEC with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "TilQazyna/Til-1B-multilingual-base-GEC" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TilQazyna/Til-1B-multilingual-base-GEC",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "TilQazyna/Til-1B-multilingual-base-GEC" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TilQazyna/Til-1B-multilingual-base-GEC",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use TilQazyna/Til-1B-multilingual-base-GEC with Docker Model Runner:
```
docker model run hf.co/TilQazyna/Til-1B-multilingual-base-GEC
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Til-1B-multilingual-base-GEC

Kazakh grammatical error correction (GEC) model, fine-tuned from Til-1B-multilingual-base — the token-balanced 941M multilingual base.

Given a Kazakh sentence that may contain spelling/grammar errors, the model returns the corrected sentence.

Task format

Plain seq2seq-as-causal-LM with a => separator (greedy decoding):

<input sentence with errors> => <corrected sentence>

Results — exact-match on the held-out novel test sets (100 items each)

Test	exact_match	chrF	keep_correct
test100	78.0	93.9	100.0
test100_v2	70.0	90.7	91.7
test100_v3	71.0	94.4	83.3
average	73.0

Context (Til GEC ladder, same recipe & test sets): this multilingual-base GEC (avg 73.0) beats the 0.5B multilingual-base GEC (avg 68.7) and closes ~40% of the gap to the best Kazakh-monolingual GEC (Til-kk-0.5B-256k-gec-exp085, avg 79.3); on the newest set (test100_v3) it matches the monolingual model (71 = 71). keep_correct ≈ 100 % on clean input means it does not over-correct already-correct text.

Training

Base: Til-1B-multilingual-base (DeepseekV3 MLA, 941M, 131 072 vocab / Til-Tokenizer-128k).
Recipe: supervised fine-tuning on "<input> => <target>" pairs, data = rule-based synthetic + 78k human-annotated + identity (keep-correct) + hard mined cases.
Hyperparameters: 3 epochs, lr 2e-5, per-device batch 16, 8 × H200, NCCL_IB_DISABLE=1.
Precision: bfloat16.

How to use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "TilQazyna/Til-1B-multilingual-base-GEC"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16).eval()

def correct(sentence):
    ids = tok(sentence + " => ", return_tensors="pt").input_ids
    out = model.generate(ids, max_new_tokens=96, do_sample=False,
                         eos_token_id=tok.eos_token_id, pad_token_id=1)
    text = tok.decode(out[0], skip_special_tokens=True)
    return text.split(" => ", 1)[-1].strip()

print(correct("мен кітап оқыдым керек"))

Limitations

Trained for sentence-level Kazakh GEC only; not a chat model. Hard/rare error types remain the weakest (~11–15/24 on the hard bucket). 4096-token context.

Model family

Part of the Til family by TilQazyna. Base: Til-1B-multilingual-base; monolingual GEC sibling: Til-kk-0.5B-256k-gec-exp085.

Downloads last month: -

Safetensors

Model size

0.9B params

Tensor type

BF16