Training and Custom Models
Create specialized models for your domain and stack without complex infrastructure: efficient fine-tuning (PEFT), adapters and merge, with reproducible evaluation. Register and deploy in one click.
"Verified compatibility" indicates internal testing until 09/2025. Some advanced functions appear as In development or Planned.
Why it's different
- Production-oriented: models by domain and by stack (Technology/Domain Packs), not generic.
- Real efficiency: PEFT (LoRA/QLoRA) to reduce cost/VRAM and accelerate iteration.
- Automatic evaluation before deployment: quality, security, robustness, latency and cost.
- Complete traceability: datasets, parameters, weights and versions with MLflow.
Training modes
Training Workflow
Guided process from data to deployed model
Training packs
Technology Packs
models fine-tuned to stacks and versions (frameworks, libraries and conventions).
Domain Packs
models fine-tuned to verticals (support, banking, retail, health) with specific metrics and evaluations.
Guided flow (from data to production)
Data
load your dataset (QA/code/tickets/manuals); validations and balancing.
Configuration
choose base and method (LoRA/QLoRA/Adapter).
Job
launch training with hardware presets (GPU/CPU); monitoring and logs.
Evaluation
run the suite (quality, security/robustness, latency/cost).
Registration
save weights and metadata in MLflow; version.
Deployment
one click to Serving module; define policies and metrics.
Automatic evaluation (reproducible)
General LLM
exact match, F1/accuracy, BLEU/ROUGE/BERTScore (according to task).
Code
compiles/"test passes", pass@k.
RAG
recall@k, MRR/nDCG, groundedness/% response with source.
Security and robustness
jailbreak/PII/toxicity, prompt injection, adversarial variations.
Operation
P50/P95 latency, throughput, cost per query/1k tokens.
Evaluation adapters to recognized methodologies; we add new ones in 48–72h if you need them.
Hardware and backends (verified compatibility)
- GA: CPU and NVIDIA (vLLM/Transformers/ONNX Runtime), OpenVINO (Intel CPU).
- Beta: AMD ROCm, Intel iGPU/NPU, Habana.
- Presets by memory/speed (balanced, performance, memory_efficient).
Key metrics you'll see
- Improvement vs. base in target metric (e.g., +F1, +exact match, +pass@k).
- Latency/cost per query compared (before/after).
- Adapter size/VRAM and total training time.
- Global health score (quality+security+operation).
Availability
Available today (v0.10 – 10/10/2025)
- PEFT (LoRA/QLoRA) and Adapters with guided flow and hardware presets.
- Basic evaluation (quality/operation) and MLflow registration.
- 1-click deployment to Serving module.
In development (v0.11 – Q4 2025)
- Expanded suites (security/robustness), comparative dashboards and more Packs.
Planned (2026)
- Advanced Model Merge, distributed multi-node jobs and expanded sector templates.