GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
concretejungles
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
concretejungles
Model Tree
Vikhrmodels/Vikhr-Qwen-2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Замеры на ru-llm-arena-hard (судья — GPT-4o, базовая модель — GPT-3.5-turbo) и MERA. Колонка Avg #Tokens — средняя длина ответа на ru-llm-arena: контроль длины удерживает её низкой при росте качества.
Evaluated on ru-llm-arena-hard (judge: GPT-4o, baseline: GPT-3.5-turbo) and MERA. The Avg #Tokens column is the mean response length on ru-llm-arena — length control keeps it low while quality improves.
| Model | ru-llm-arena-hard | MERA | Avg #Tokens |
|---|---|---|---|
| Qwen/Qwen2.5-3B-Instruct | 38.1 | 0.423 | 431 |
| Qwen2.5-1.5B-Instruct-GRPO-LC | 36.1 | 0.301 | 558 |
| Vikhrmodels/QVikhr-2.5-1.5B-Instruct-SMPO | 33.4 | 0.301 | 510 |
| Qwen/Qwen3-1.7B (non-thinking) | 31.4 | 0.3 | 910 |
Базовая SFT-модель Vikhr-Qwen-2.5-1.5B-Instruct дообучена онлайн-методом GRPO (вариант Dr. GRPO, scale_rewards=none). Для каждого промпта генерировалось 16 ответов, которые оценивались reward-моделью Skywork-Reward-V2-Llama-3.1-8B. Поскольку reward-модель смещена в сторону длинных ответов, к награде добавлялся штраф за длину (целевая длина ≈ 640 токенов) — это удерживает среднюю длину ответа низкой и борется с «раздуванием». Промпты — подвыборка датасета t-tech/T-Wix, отобранная по разнообразию инструкций с помощью InsTag.
Ключевые гиперпараметры: beta=0.1, learning_rate=1e-6, косинусный планировщик с прогревом 10%, 1 эпоха, temperature=0.9, max_completion_length=2048, обучение в bf16. Генерация на vLLM (colocate).
The Vikhr-Qwen-2.5-1.5B-Instruct SFT base was further trained with the online RL method GRPO (Dr. GRPO variant, scale_rewards=none). For each prompt, 16 candidate responses were sampled and scored by the Skywork-Reward-V2-Llama-3.1-8B reward model. Because the reward model is length-biased toward longer answers, a length penalty was added to the reward (target length ≈ 640 tokens) to keep mean response length low and counteract verbosity inflation. Prompts were an instruction-diversity (InsTag) subsample of the t-tech/T-Wix dataset.
Key hyperparameters: beta=0.1, learning_rate=1e-6, cosine schedule with 10% warmup, 1 epoch, temperature=0.9, max_completion_length=2048, bf16 training. Generation served by vLLM (colocate mode).
Рекомендуемые параметры генерации совпадают с generation_config.json модели / Recommended generation settings match the model's generation_config.json (temperature=0.7, top_p=0.8, top_k=20, repetition_penalty=1.1).
python
from transformers import AutoModelForCausalLM, AutoTokenizer # Load the model and tokenizermodel_name = "concretejungles/Qwen2.5-1.5B-Instruct-GRPO-LC"model = AutoModelForCausalLM.from_pretrained(model_name)tokenizer = AutoTokenizer.from_pretrained(model_name) # Prepare the input textinput_text = "Напиши краткое описание книги Гарри Поттер." messages = [ {"role": "system", "content": "Вы — ИИ-помощник. Отвечайте полезно, честно и безопасно."}, {"role": "user", "content": input_text},] # Tokenize and generate textinput_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")output = model.generate( input_ids, max_new_tokens=1024, do_sample=True, temperature=0.7, top_p=0.8, top_k=20, repetition_penalty=1.1,) # Decode and print only the newly generated tokensgenerated_text = tokenizer.decode(output[0][input_ids.shape[-1]:], skip_special_tokens=True)print(generated_text)Результат был получен во время работы в СКБ Контур.
Основано на Vikhrmodels/Vikhr-Qwen-2.5-1.5B-Instruct.
text
@inproceedings{nikolich2024vikhr, title={Vikhr: Advancing Open-Source Bilingual Instruction-Following Large Language Models for Russian and English}, author={Aleksandr Nikolich and Konstantin Korolev and Sergei Bratchikov and Nikolay Kompanets and Igor Kiselev and Artem Shelmanov}, booktitle={Proceedings of the 4th Workshop on Multilingual Representation Learning (MRL) @ EMNLP-2024}, year={2024}, publisher={Association for Computational Linguistics}, url={https://arxiv.org/pdf/2405.13929}}| Vikhrmodels/Vikhr-Qwen-2.5-1.5B-Instruct | 29.8 | 0.295 | 636 |
| RefalMachine/RuadaptQwen2.5-1.5B-instruct | 20.1 | 0.364 | 549 |
| Qwen/Qwen2.5-1.5B-Instruct | 11.3 | 0.357 | 362 |