Instructions to use AuroraSystem/Velary-2.0-1.7B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use AuroraSystem/Velary-2.0-1.7B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="AuroraSystem/Velary-2.0-1.7B", trust_remote_code=True) messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("AuroraSystem/Velary-2.0-1.7B", trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained("AuroraSystem/Velary-2.0-1.7B", trust_remote_code=True, device_map="auto") messages = [ {"role": "user", "content": "Who are you?"}, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use AuroraSystem/Velary-2.0-1.7B with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "AuroraSystem/Velary-2.0-1.7B" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "AuroraSystem/Velary-2.0-1.7B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/AuroraSystem/Velary-2.0-1.7B
- SGLang
How to use AuroraSystem/Velary-2.0-1.7B 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 "AuroraSystem/Velary-2.0-1.7B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "AuroraSystem/Velary-2.0-1.7B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'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 "AuroraSystem/Velary-2.0-1.7B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "AuroraSystem/Velary-2.0-1.7B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use AuroraSystem/Velary-2.0-1.7B with Docker Model Runner:
docker model run hf.co/AuroraSystem/Velary-2.0-1.7B
Velary 2.0
English
Velary 2.0 is a 1.7B-parameter bilingual (Russian/English) instruction model built on
LiquidAI/LFM3-1.7B (LFM3 tri-hybrid architecture, 256K-token context) with a reasoning
LoRA merged into the base weights.
Benchmarks (ABench75, 375 tasks, automatic regex scoring)
| Split | Tasks | Score |
|---|---|---|
| Public | 300 | 92.0% |
| Private | 75 | 93.3% |
| Combined | 375 | 92.3% |
Measured with the high reasoning-effort system prompt, greedy decoding and
max_new_tokens=384. Only the overall score is reported.
Training details
- Base model:
LiquidAI/LFM3-1.7B(Lfm3ForCausalLM, 24 layers, hidden 2048, vocab 128000, tri-hybrid convolution/attention stack). - Method: LoRA, r=16, alpha=32, dropout 0.05, merged into the base weights with PEFT. No adapter files are needed at inference time.
- Data: ~14k synthetic/teacher-generated reasoning examples (14,086 train / 322 eval),
ChatML format with
thinkingtraces. - Precision: bfloat16. Trained at 2048-token sequences; the base supports 256K but quality beyond the training length is not validated.
Reasoning control
Reasoning depth is selected through the system prompt:
You are Aurora, a helpful AI assistant created by Aurora System. Reasoning effort: high. Think carefully and thoroughly, then give a complete structured answer.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("AuroraSystem/Velary-2.0-1.7B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
"AuroraSystem/Velary-2.0-1.7B", trust_remote_code=True,
dtype="bfloat16", device_map="auto")
messages = [
{"role": "system", "content": "You are Aurora, a helpful AI assistant created by Aurora System. "
"Reasoning effort: high. Think carefully and thoroughly, then give a "
"complete structured answer."},
{"role": "user", "content": question},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True,
return_tensors="pt", return_dict=True)
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
Limitations
- Trained on synthetic reasoning data: verify facts and Russian orthography before use.
- Quality is validated at 2048-token sequences; longer contexts are untested.
- Casual conversation and open-ended creative writing are limited.
- Longer reasoning increases token usage; use
max_new_tokensof at least 512 for high effort. - The model is not safety-aligned beyond the training data.
ะ ัััะบะธะน
Velary 2.0 โ ะดะฒััะทััะฝะฐั (ััััะบะธะน/ะฐะฝะณะปะธะนัะบะธะน) ะธะฝััััะบัะธะฒะฝะฐั ะผะพะดะตะปั ะฝะฐ 1.7B ะฟะฐัะฐะผะตััะพะฒ
ะฝะฐ ะฑะฐะทะต LiquidAI/LFM3-1.7B (ััะธ-ะณะธะฑัะธะดะฝะฐั ะฐัั
ะธัะตะบัััะฐ LFM3, ะบะพะฝัะตะบัั 256K) ั reasoning-LoRA,
ัะผะตัะดะถะตะฝะฝะพะน ะฒ ะฑะฐะทะพะฒัะต ะฒะตัะฐ.
ะะตะฝัะผะฐัะบะธ (ABench75, 375 ะทะฐะดะฐั, ะฐะฒัะพะผะฐัะธัะตัะบะฐั ะฟัะพะฒะตัะบะฐ ะฟะพ ัะตะณัะปััะบะต)
| ะกะฟะปะธั | ะะฐะดะฐั | ะะฐะปะป |
|---|---|---|
| Public | 300 | 92.0% |
| Private | 75 | 93.3% |
| ะัะตะณะพ | 375 | 92.3% |
ะะฐะผะตั ั ัะธััะตะผะฝัะผ ะฟัะพะผัะพะผ ะฒััะพะบะพะณะพ ััะพะฒะฝั ัะฐัััะถะดะตะฝะธะน, greedy-ะดะตะบะพะดะธัะพะฒะฐะฝะธะตะผ ะธ
max_new_tokens=384. ะะพะบะฐะทะฐะฝ ัะพะปัะบะพ ะพะฑัะธะน ะฑะฐะปะป.
ะะตัะฐะปะธ ะพะฑััะตะฝะธั
- ะะฐะทะพะฒะฐั ะผะพะดะตะปั:
LiquidAI/LFM3-1.7B(Lfm3ForCausalLM, 24 ัะปะพั, hidden 2048, ัะปะพะฒะฐัั 128000, ะณะธะฑัะธะดะฝัะน ััะตะบ ัะฒัััะพะบ ะธ ะฒะฝะธะผะฐะฝะธั). - ะะตัะพะด: LoRA, r=16, alpha=32, dropout 0.05, ั ะฟะพัะปะตะดัััะธะผ ะผะตัะดะถะตะผ ะฒ ะฑะฐะทะพะฒัะต ะฒะตัะฐ ัะตัะตะท PEFT. ะะดะฐะฟัะตัั ะฟัะธ ะธะฝัะตัะตะฝัะต ะฝะต ะฝัะถะฝั.
- ะะฐะฝะฝัะต: ~14k ัะธะฝัะตัะธัะตัะบะธั
/ััะธัะตะปััะบะธั
ะฟัะธะผะตัะพะฒ ัะฐัััะถะดะตะฝะธะน (14 086 train / 322 eval),
ัะพัะผะฐั ChatML ั ััะตะนัะฐะผะธ
thinking. - ะขะพัะฝะพััั: bfloat16. ะะฑััะตะฝะธะต ะฝะฐ ะฟะพัะปะตะดะพะฒะฐัะตะปัะฝะพัััั 2048 ัะพะบะตะฝะพะฒ; ะฑะฐะทะฐ ะดะตัะถะธั 256K, ะฝะพ ะบะฐัะตััะฒะพ ะทะฐ ะฟัะตะดะตะปะฐะผะธ ะพะฑััะฐััะตะน ะดะปะธะฝั ะฝะต ะฟัะพะฒะตััะปะพัั.
ะฃะฟัะฐะฒะปะตะฝะธะต ัะฐัััะถะดะตะฝะธัะผะธ
ะะปัะฑะธะฝะฐ ะทะฐะดะฐัััั ัะธััะตะผะฝัะผ ะฟัะพะผัะพะผ:
You are Aurora, a helpful AI assistant created by Aurora System. Reasoning effort: high. Think carefully and thoroughly, then give a complete structured answer.
ะะณัะฐะฝะธัะตะฝะธั
- ะะฐัะตััะฒะพ ะฟะพะดัะฒะตัะถะดะตะฝะพ ะฝะฐ 2048 ัะพะบะตะฝะฐั ; ะดะปะธะฝะฝัะต ะบะพะฝัะตะบััั ะฝะต ัะตััะธัะพะฒะฐะปะธัั.
- ะกะปะฐะฑะพ ัะฐะทะฒะธัะฐ ัะฒะตััะบะฐั ะฑะตัะตะดะฐ ะธ ะพัะบัััะพะต ะบัะตะฐัะธะฒะฝะพะต ะฟะธััะผะพ.
- ะะปะธะฝะฝัะต ัะฐัััะถะดะตะฝะธั ัะฒะตะปะธัะธะฒะฐัั ัะฐัั
ะพะด ัะพะบะตะฝะพะฒ; ะดะปั high ััะฐะฒััะต
max_new_tokensะฝะต ะผะตะฝััะต 512. - ะะพะดะตะปั ะฝะต ะฒััะพะฒะฝะตะฝะฐ ะฟะพ ะฑะตะทะพะฟะฐัะฝะพััะธ ัะฒะตัั ะพะฑััะฐััะธั ะดะฐะฝะฝัั .
- Downloads last month
- -