Trendyol

Trendyol/Trendyol-Cybersecurity-LLM-v2-70B-Q4KM (Llama‑3.3‑70B Finetuned) > v2-Max is a defense-focused, alignment-safe cybersecurity language model based on Llama-3.3-70B. It was trained via SFT from scratch on the v2 dataset and ranked near the top on CS-Eval in the EN and EN-ZH tracks. Developed by the Trendyol Group Security Team. Developed Trendyol Group Security Team - Alican Kiraz - İsmail Yavuz - Melih Yılmaz - Mertcan Kondur - Rıza Sabuncu - Özgün Kultekin We thank Ahmet Gökhan Yalçın, Cenk Çivici, Nezir Alp, Yiğit Darçın for all their support. - English: 3rd place - English–Chinese: 5th place - Comprehensive average: 91.03 | Category | Score | |---|---:| | Fundamentals of System Security and Software Security | 91.33 | | Access Control and Identity Management | 90.23 | | Encryption Technology and Key Management | 92.70 | | Infrastructure Security | 91.85 | | AI and Network Security | 94.55 | | Vulnerability Management and Penetration Testing | 90.78 | | Threat Detection and Prevention | 92.44 | | Data Security and Privacy Protection | 90.48 | | Supply Chain Security | 93.69 | | Security Architecture Design | 87.80 | | Business Continuity & Emergency Response / Recovery | 85.00 | | Chinese Task | 91.07 | Differences: Old model (Qwen3-14B, “BaronLLM v2.0”) → New model (Llama-3.3-70B, “v2-Max”): Base model leap: 14B → 70B (stronger reasoning, long-range context, and standards-compliant answers). Dataset: v1.1 (21,258) → v2.0 (83,920) rows (≈4×); coverage includes OWASP, MITRE ATT&CK, NIST CSF, CIS, ASD Essential 8, modern identity (OAuth2/OIDC/SAML), TLS, Cloud & DevSecOps, Cryptography, and AI Security. Security gates: adversarial refusal tests against jailbreak/prompt injection, static policy linting, content risk taxonomy; refuse-or-report strategies in system prompts. Orientation: offensive examples were removed; defensive, safe outputs are preferred. Size: 83,920 system/user/assistant triplets License: Apache-2.0 Language: English Split: `train` (100%) Format: Parquet (columns: `system`, `user`, `assistant`) Quality Gates: Deduplication, PII scrubbing, hallucination screening, adversarial refusal tests, static policy linting, risk taxonomy, strict schema validations (stable row IDs). OWASP Top 10, MITRE ATT&CK, NIST CSF, CIS Controls, ASD Essential 8 Modern identity: OAuth 2.0 / OIDC / SAML SSL/TLS practices and certificate hygiene Cloud & DevSecOps: IAM, secrets management, CI/CD, container/K8s hardening, logging/SIEM, IR runbooks Cryptography implementation hygiene AI Security (prompt injection, data poisoning, model/embedding security, etc.) > The dataset is commercial-friendly: Apache-2.0. Harmful/exploitative content and high-risk materials such as raw shellcode were excluded during dataset construction; patterns for generating safe alternatives to harmful requests were included. Defense-Focused Reasoning: Standards-aligned (OWASP/ATT&CK/NIST/CIS) recommendations with step-by-step explanations that include the “why/evidence.” Policy & Architecture Guides: Identity/access, encryption, network segmentation, cloud control sets, data classification, and alert logic. IR & Threat Hunting Support: Incident flow, triage checklists, safe log query patterns, playbook skeletons. Cloud & DevSecOps: CI/CD security gates, IaC misconfiguration patterns, K8s hardening checklists. Refusal-by-Design: Safe alternatives and compliance-aligned response patterns for exploitative/malicious prompts. > Note: This model is SFT-based; it has no tool use, web browsing, or access to an executable environment. In expert operations it is an assistant; it does not, by itself, constitute evidence. GGUF / llama.cpp > For practical single-machine use with a 70B model, Q4KM is recommended; higher-quality GGUF variants require much more memory. | Goal | Template | Note | | --------------------- | ------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------- | | Risk Assessment | `ROLE: Security Architect\nCONTEXT: ...\nTASK: Produce a control-by-control gap analysis against NIST CSF v2.0...` | `temperature=0.2–0.4, topp=0.9` | | IR Playbook | `Create a phase-by-phase incident response playbook for ransomware in hybrid (Azure+on-prem) AD.` | Ask for a validation checklist in the final step. | | Identity Security | `Propose hardened OAuth2/OIDC configs for a multi-tenant SPA+API.` | Request an "abuse cases" list at the end of the response. | | Cloud Guardrails | `Generate K8s hardening controls mapped to CIS + NSA/CISA.` | Include “rationale” and “validation” columns. | Intended Use: Enterprise defense consulting, architectural guidance, control mappings, IR/TI support outputs, training, and documentation. Unauthorized penetration testing, PoC exploit/payload generation, instructions that harm live systems. Training or enriching outputs with personal data or confidential information. Explicit refusal for harmful requests + safe alternative suggestions. Content policies are tested within the dataset (against jailbreak/prompt injection). Disclaimer: Model outputs should not be applied automatically without expert review; corporate policy and regulation take precedence. Training Method: SFT (instruction tuning) + system-prompt guardrails. Context Window: Llama-3.3-70B’s default (same as Meta’s release). Languages: Training language is English; EN-ZH evaluation is supported. Release: Weights under the Meta Llama 3.3 License; dataset under Apache-2.0. Issues / Feedback: HF or GitHub Issues Research Collaboration: Trendyol Group Security Team Acknowledgments: Trendyol Security Team, community contributors, and independent evaluators 2025-10-06 — v2-Max: Migration to the Llama-3.3-70B base; v2 dataset with 83,920 rows; alignment security gates; CS-Eval EN #3 / EN-ZH #5, average score 91.03.

Trendyol/TY-ecomm-embed-multilingual-base-v1.2.0 is a multilingual sentence-transformers embedding model fine-tuned on e-commerce datasets, optimized for semantic similarity, search, classification, and retrieval tasks. It is integrating domain-specific signals from millions of real-world queries, product descriptions, and user interactions. This model is fine-tuned over a distilled version of Alibaba-NLP/gte-multilingual-base using Turkish-English pair translation dataset. Keynotes: Optimized for e-commerce semantic search Enhanced Turkish and multilingual query understanding Supports query rephrasing and paraphrase mining Robust for product tagging and attribute extraction Suitable for clustering and product categorization High-performance in semantic textual similarity 384-token input support 768-dimensional dense vector outputs Built-in cosine similarity for inference Model Description - Model Type: Sentence Transformer - Maximum Sequence Length: 384 tokens - Output Dimensionality: 768 dimensions - Matryoshka Dimensions: 768, 512, 128 - Similarity Function: Cosine Similarity - Training Datasets: - Multilingual and Turkish search terms - Turkish instruction datasets - Turkish summarization datasets - Turkish e-commerce rephrase datasets - Turkish question-answer pairs - and more! - Documentation: Sentence Transformers Documentation - Repository: Sentence Transformers on GitHub - Hugging Face: Sentence Transformers on Hugging Face While this model is trained on e-commerce-related datasets, including multilingual and Turkish data, users should be aware of several limitations: Domain bias: Performance may degrade for content outside the e-commerce or product-related domains, such as legal, medical, or highly technical texts. Language coverage: Although multilingual data was included, majority of the dataset is created in Turkish. Input length limitations: Inputs exceeding the maximum sequence length (384 tokens) will be truncated, potentially losing critical context in long texts. Spurious similarity: Semantic similarity may incorrectly assign high similarity scores to unrelated but lexically similar or frequently co-occurring phrases in training data. Human Oversight: We recommend incorporating a human curation layer or using filters to manage and improve the quality of outputs, especially in public-facing applications. This approach can help mitigate the risk of generating objectionable content unexpectedly. Application-Specific Testing: Developers intending to use Trendyol embedding models should conduct thorough safety testing and optimization tailored to their specific applications. This is crucial, as the model’s outputs may occasionally be biased or inaccurate. Responsible Development and Deployment: It is the responsibility of developers and users of Trendyol embedding models to ensure its ethical and safe application. We urge users to be mindful of the model's limitations and to employ appropriate safeguards to prevent misuse or harmful consequences. Training Hyperparameters Non-Default Hyperparameters - `overwriteoutputdir`: True - `evalstrategy`: steps - `perdevicetrainbatchsize`: 2048 - `perdeviceevalbatchsize`: 128 - `learningrate`: 0.0005 - `numtrainepochs`: 1 - `warmupratio`: 0.01 - `fp16`: True - `ddptimeout`: 300000 - `batchsampler`: noduplicates - `overwriteoutputdir`: True - `dopredict`: False - `evalstrategy`: steps - `predictionlossonly`: True - `perdevicetrainbatchsize`: 2048 - `perdeviceevalbatchsize`: 128 - `pergputrainbatchsize`: None - `pergpuevalbatchsize`: None - `gradientaccumulationsteps`: 1 - `evalaccumulationsteps`: None - `torchemptycachesteps`: None - `learningrate`: 0.0005 - `weightdecay`: 0.0 - `adambeta1`: 0.9 - `adambeta2`: 0.999 - `adamepsilon`: 1e-08 - `maxgradnorm`: 1.0 - `numtrainepochs`: 1 - `maxsteps`: -1 - `lrschedulertype`: linear - `lrschedulerkwargs`: {} - `warmupratio`: 0.01 - `warmupsteps`: 0 - `loglevel`: passive - `loglevelreplica`: warning - `logoneachnode`: True - `loggingnaninffilter`: True - `savesafetensors`: True - `saveoneachnode`: False - `saveonlymodel`: False - `restorecallbackstatesfromcheckpoint`: False - `nocuda`: False - `usecpu`: False - `usempsdevice`: False - `seed`: 42 - `dataseed`: None - `jitmodeeval`: False - `useipex`: False - `bf16`: False - `fp16`: True - `fp16optlevel`: O1 - `halfprecisionbackend`: auto - `bf16fulleval`: False - `fp16fulleval`: False - `tf32`: None - `localrank`: 0 - `ddpbackend`: None - `tpunumcores`: None - `tpumetricsdebug`: False - `debug`: [] - `dataloaderdroplast`: True - `dataloadernumworkers`: 0 - `dataloaderprefetchfactor`: None - `pastindex`: -1 - `disabletqdm`: False - `removeunusedcolumns`: True - `labelnames`: None - `loadbestmodelatend`: False - `ignoredataskip`: False - `fsdp`: [] - `fsdpminnumparams`: 0 - `fsdpconfig`: {'minnumparams': 0, 'xla': False, 'xlafsdpv2': False, 'xlafsdpgradckpt': False} - `fsdptransformerlayerclstowrap`: None - `acceleratorconfig`: {'splitbatches': False, 'dispatchbatches': None, 'evenbatches': True, 'useseedablesampler': True, 'nonblocking': False, 'gradientaccumulationkwargs': None} - `deepspeed`: None - `labelsmoothingfactor`: 0.0 - `optim`: adamwtorch - `optimargs`: None - `adafactor`: False - `groupbylength`: False - `lengthcolumnname`: length - `ddpfindunusedparameters`: None - `ddpbucketcapmb`: None - `ddpbroadcastbuffers`: False - `dataloaderpinmemory`: True - `dataloaderpersistentworkers`: False - `skipmemorymetrics`: True - `uselegacypredictionloop`: False - `pushtohub`: False - `resumefromcheckpoint`: None - `hubmodelid`: None - `hubstrategy`: everysave - `hubprivaterepo`: None - `hubalwayspush`: False - `gradientcheckpointing`: False - `gradientcheckpointingkwargs`: None - `includeinputsformetrics`: False - `includeformetrics`: [] - `evaldoconcatbatches`: True - `fp16backend`: auto - `pushtohubmodelid`: None - `pushtohuborganization`: None - `mpparameters`: - `autofindbatchsize`: False - `fulldeterminism`: False - `torchdynamo`: None - `rayscope`: last - `ddptimeout`: 300000 - `torchcompile`: False - `torchcompilebackend`: None - `torchcompilemode`: None - `dispatchbatches`: None - `splitbatches`: None - `includetokenspersecond`: False - `includenuminputtokensseen`: False - `neftunenoisealpha`: None - `optimtargetmodules`: None - `batchevalmetrics`: False - `evalonstart`: False - `useligerkernel`: False - `evalusegatherobject`: False - `averagetokensacrossdevices`: False - `prompts`: None - `batchsampler`: noduplicates - `multidatasetbatchsampler`: proportional Framework Versions - Python: 3.11.11 - Sentence Transformers: 3.4.1 - Transformers: 4.48.1 - PyTorch: 2.5.1+cu124 - Accelerate: 1.5.1 - Datasets: 2.21.0 - Tokenizers: 0.21.1

Trendyol-Cybersecurity-LLM-Qwen3-32B-Q80-GGUF represents a paradigmatic shift in the application of large language models to the cybersecurity domain. This model, architected upon the Qwen3-32B foundation and optimized through Q80 quantization in GGUF format, embodies a sophisticated approach to AI-driven security operations. The model's development reflects a comprehensive understanding of the intricate requirements of modern cybersecurity practices, integrating advanced natural language processing capabilities with domain-specific expertise. - Architecture: Qwen3-32B base model with specialized cybersecurity fine-tuning utilizing advanced transformer architectures - Quantization: Q80 GGUF format implementing optimal performance-to-precision trade-offs - Training Infrastructure: 3×NVIDIA H200 GPUs with distributed training paradigms - Training Duration: ~100 hours (approximately 2 months of iterative training with continuous evaluation) - Non-commercial: This model operates under strict non-profit principles - Safety-first Design: Incorporates multi-layered safety mechanisms to prevent malicious exploitation The model demonstrates exceptional proficiency across six critical cybersecurity verticals, each representing a distinct operational paradigm within the security ecosystem: 1. Incident Response (IR) Advanced capabilities in orchestrating comprehensive incident response workflows: 2. Threat Hunting Proactive threat detection utilizing advanced behavioral analytics: 3. Code Analysis Multi-paradigm code security assessment framework: 4. Exploit Development Ethical exploit engineering for security validation: 5. Reverse Engineering Advanced binary and protocol analysis capabilities: 6. Malware Analysis Sophisticated malware examination and classification system: We welcome contributions from the global cybersecurity community. Our contribution framework ensures high-quality, security-focused enhancements: For academic and research collaborations, please refer to our research guidelines and dataset access protocols. We maintain partnerships with leading cybersecurity research institutions and welcome new collaborative opportunities. This model is released under the Apache 2.0 License with additional ethical use provisions specific to cybersecurity applications. 🛡️ Developed with Passion by Trendyol Security Team 🛡️ Empowering the cybersecurity community with advanced AI capabilities Together, we build a more secure digital future

Background Removal is an IS-Net–based human segmentation and background-removal model designed to automatically detect and isolate people in images. It produces high-quality binary/alpha masks and trimmed RGBA composites intended for downstream editing, compositing, and automated image pipelines. Although optimized for fashion photography, it is suitable for any application where the image contains human and the goal is to separate them cleanly from the background. - Architecture: IS-Net - Objective: Fine-tuning isnet-general-use model with TY fashion images to better performance of fashion images - Training Data: Large-scale Trendyol fashion product image dataset containing human models - Hardware: Multi-GPU training with PyTorch - Framework: PyTorch - Automatically remove backgrounds from images containing human, isolating the subject for further editing, compositing, or analysis. - Designed for use in applications such as e-commerce product photography, fashion catalogs, profile pictures, and creative media projects where the human subject needs to be cleanly separated from the background. - Optimized for images with clear human presence; not intended for objects, animals, or scenes without people. - Can be used as a preprocessing step for downstream tasks like virtual try-on, background replacement, and image-based content generation. Complete example to load the model, remove background of an image, and save the results: - Achieve high-accuracy image matting: Especially for intricate details on human models, such as hair and clothing textures. - Backbone: IS-Net general use model trained on DIS dataset V1.0: DIS5K - Model Input Size: 1800x1200 - Training Framework: Torch 1.13.1 - Domain Specificity: Optimized for e-commerce fashion product images with human models included; may not generalize well to other image domains - Image Quality: Performance may degrade on low-quality, heavily compressed, or significantly distorted images - Category Bias: Performance may vary across different product categories based on training data distribution - Commercial Use: Designed for e-commerce applications; consider potential impacts on market competition - Privacy: Ensure compliance with data protection regulations when processing product images - Fairness: Monitor for biased similarity judgments across different product categories or brands This model is released by Trendyol as a source-available, non-open-source model. - View, download, and evaluate the model weights. - Use the model for non-commercial research and internal testing. - Use the model or its derivatives for commercial purposes, provided that: - You cite Trendyol as the original model creator. - You notify Trendyol in advance via [email protected] or other designated contact. - Redistribute or host the model or its derivatives on third-party platforms without prior written consent from Trendyol. - Use the model in applications violating ethical standards, including but not limited to surveillance, misinformation, or harm to individuals or groups. By downloading or using this model, you agree to the terms above. For technical support or questions about this model, please contact the Trendyol Data Science team.