--- title: Spark-Scilit-X1-13B canonical_url: "https://www.modelscope.cn/models/iflytek/Spark-Scilit-X1-13B" md_url: "https://www.modelscope.cn/models/iflytek/Spark-Scilit-X1-13B.md" repository: iflytek/Spark-Scilit-X1-13B chinese_name: Spark-Scilit-X1-13B last_updated: 2025-10-20 license: "Apache License 2.0" pipeline_tag: text-generation tasks: - text-generation model_type: - spark architectures: - SparkModel parameters: 13.3B tensor_type: - F32 library_name: - transformer - safetensors - pytorch frameworks: - PyTorch language: - en - zh downloads: 24519 stars: 8 tags: - CoT - "scientific llm" - reasoning --- # Spark-Scilit-X1-13B > Spark-Scilit-X1-13B - iflytek 在 ModelScope 开源的模型。iFLYTEK Spark Scilit-X1-13B is based on the latest generation of iFLYTEK Foundation Model, and has been trained on multiple core tasks derived from scientific literature. As a large language model tailored for… iflytek/Spark-Scilit-X1-13B 是 ModelScope 魔搭社区上的 13.3B 参数text-generation模型，采用 Apache License 2.0 许可。 - **Repository**: iflytek/Spark-Scilit-X1-13B - **License**: Apache License 2.0 - **Tasks**: text-generation - **Parameters**: 13.3B - **Tags**: CoT, scientific llm, reasoning - **Downloads**: 24519 - **Stars**: 8 - **Last updated**: 2025-10-20 Source: https://www.modelscope.cn/models/iflytek/Spark-Scilit-X1-13B --- # iFLYTEK Spark Scilit-X1-13B Model Card ## Model Introduction iFLYTEK Spark Scilit-X1-13B is based on the latest generation of iFLYTEK Foundation Model, and has been trained on multiple core tasks derived from scientific literature. As a large language model tailored for academic research scenarios, it has shown excellent performance in Paper Assisted Reading, Academic Translation, English Polishing, and Review Generation, aiming to provide efficient and accurate intelligent assistance for researchers, faculty members, and students. **Key Features** - **Deep Reasoning Architecture**: A unified framework that combines long Chain-of-Thought (CoT) reasoning with dual-process theory, supporting both fast (reactive) and slow (deliberative) thinking modes. - **Domain Knowledge Optimization**: Trained on large-scale, high-quality scientific corpora, the model shows strong capability in understanding domain-specific terminology and generating academically appropriate expressions. - **Balancing Scientific Research Ability and Universality**: By adopting multi-stage training techniques, the model maintains strong scientific research ability while also exhibiting robust general task-solving performance. ## Model Summary | Parameter | Value | |--------------------------|----------------| | Total Parameters | 13B | | Context Length | 32K | | Window Length | 32K | | Number of Layers | 40 | | Attention Hidden Dim | 5120 | | Attention Heads | 40 | | Vocabulary Size | 130K | | Attention Mechanism | GQA | | Activation Function | GeLU | *Model specifications for the iFLYTEK Spark Scilit-X1-13B architecture.* ## Benchmark Results | Task | Metric | Spark-Scilit-X1-13B | Qwen3-32B | Qwen3-Next-80B-A3B | DeepSeek-R1 | O3 | |---------------------------------|--------|---------------------|-----------|--------------------|-------------|------| | Paper Assisted Reading | MOS | 4.04 | 3.98 | 4.06 | 4.01 | 4.1 | | Academic Translation | MOS | 4.18 | 4.04 | 4.08 | 4.12 | 4.22 | | English Polishing | MOS | 4.22 | 4.11 | 4.2 | 3.98 | 4.28 | | Review Generation | MOS | 3.88 | 3.5 | 3.7 | 3.68 | 4.01 | *All metrics are Mean Opinion Scores (MOS) averaged over human evaluations (1–5 scale).* **Evaluation Notes**: 1. All baseline models were evaluated through their official APIs. 2. Spark Scilit-X1-13B was evaluated on a single NVIDIA A100 (80 GB) GPU. 3. All reported results represent averaged zero-shot performance. 4. A consistent evaluation protocol was applied across all models. ## Usage **requirement** ```bash cd /path/to/Spark-Scilit-X1-13B # We recommend using Python 3.10 pip install -r requirements.txt pip install . ``` ```python from transformers import AutoModelForCausalLM from tokenizer_spark import SparkTokenizer # Load model and tokenizer model_name = "iflytek/Spark-Scilit-X1-13B" tokenizer = SparkTokenizer.from_pretrained(model_name) model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype=torch.bfloat16, device_map="auto" ) # Reactive chat_history = [ { "role" : "system", "content" : "你能够回答用户的各种问题，回答问题能够角度全面、表述专业、重点突出。" }, { "role" : "user", "content" : "你将进行论文片段的翻译任务,请将给定论文片段翻译成中文。论文片段如下:Super-hydrophobic delivery (SHD) is an efficient approach to enrich trace analytes into hot spot regions for ultrasensitive surface-enhanced Raman scattering (SERS) detection. In this article, we propose an efficient and simple method to prepare a highly uniform SHD-SERS platform of high performance in trace detection, named as “silver-nanoparticle-grafted silicon nanocones” (termed AgNPs/SiNC) platform. It is fabricated via droplet-confined electroless deposition on the super-hydrophobic SiNC array. The AgNPs/SiNC platform allows trace analytes enriched into hot spots formed by AgNPs, leading to excellent reproducibility and sensitivity. The relative standard deviation (RSD) for detecting R6G (10⁻⁶ M) is down to 4.70%, and the lowest detection concentration for R6G is 10⁻⁸ M. Moreover, various contaminants in complex liquid environments, such as crystal violet (10⁻⁶ M) in lake water, melamine (10⁻⁶ M) in liquid milk, and methyl parathion (10⁻⁶ M) in tap water, can be detected using the SERS platform. This result demonstrates the great potential of the AgNPs/SiNC platform in the fields of food safety and environmental monitoring." }] inputs = tokenizer.apply_chat_template( chat_history, tokenize=True, return_tensors="pt", add_generation_prompt=True ).to(model.device) outputs = model.generate( inputs["input_ids"], max_new_tokens=32768, top_k=10, do_sample=True, repetition_penalty=1.1, temperature=0.5, eos_token_id=5, pad_token_id=0, output_attentions=True ) response = tokenizer.decode( outputs[0][inputs.shape[1] :], skip_special_tokens=True ) print(reponse) ``` ```python # Deliberative chat_history = [ { "role" : "system", "content" : "你能够回答用户的各种问题，回答问题能够角度全面、表述专业、重点突出。当前是慢思考模式，请你先深入剖析给出问题的关键要点与内在逻辑，生成思考过程，再根据思考过程回答给出问题。思考过程以开头，在结尾处用标注结束，后为基于思考过程的回答内容。" } , { "role" : "user", "content" : "你将进行论文片段的翻译任务,请将给定论文片段翻译成中文。论文片段如下:Super-hydrophobic delivery (SHD) is an efficient approach to enrich trace analytes into hot spot regions for ultrasensitive surface-enhanced Raman scattering (SERS) detection. In this article, we propose an efficient and simple method to prepare a highly uniform SHD-SERS platform of high performance in trace detection, named as “silver-nanoparticle-grafted silicon nanocones” (termed AgNPs/SiNC) platform. It is fabricated via droplet-confined electroless deposition on the super-hydrophobic SiNC array. The AgNPs/SiNC platform allows trace analytes enriched into hot spots formed by AgNPs, leading to excellent reproducibility and sensitivity. The relative standard deviation (RSD) for detecting R6G (10⁻⁶ M) is down to 4.70%, and the lowest detection concentration for R6G is 10⁻⁸ M. Moreover, various contaminants in complex liquid environments, such as crystal violet (10⁻⁶ M) in lake water, melamine (10⁻⁶ M) in liquid milk, and methyl parathion (10⁻⁶ M) in tap water, can be detected using the SERS platform. This result demonstrates the great potential of the AgNPs/SiNC platform in the fields of food safety and environmental monitoring." }] inputs = tokenizer.apply_chat_template( chat_history, tokenize=True, return_tensors="pt", add_generation_prompt=True ).to(model.device) outputs = model.generate( inputs["input_ids"], max_new_tokens=32768, top_k=10, do_sample=True, repetition_penalty=1.1, temperature=0.5, eos_token_id=5, pad_token_id=0, output_attentions=True ) response = tokenizer.decode( outputs[0][inputs.shape[1] :], skip_special_tokens=True ) print(reponse) ``` ## License Agreement iFLYTEK Spark Scilit-X1-13B is licensed under Apache 2.0.