Kiel-2-Sol
Kiel-2-Sol is an enterprise-grade, high-performance conversational language model fine-tuned specifically to anchor the premium tier of the KielTech AI production ecosystem. Built upon Meta's meta-llama/Llama-3.2-3B-Instruct architecture, the model undergoes an intensive parameter-efficient adaptation process targeting both attention mechanism layers and core multi-layer perceptron (MLP) blocks.
This model balances dense reasoning capacities and highly structured instruction-following capabilities, making it ideal for robust enterprise automation, complex multi-turn API workflows, and rapid serverless deployment via engines like vLLM.
Model Details
Model Description
- Developed by: KielTech
- Model Type: Causal Language Model (Transformer Architecture)
- Base Model: meta-llama/Llama-3.2-3B-Instruct
- Language(s): English
- License: Apache 2.0
- Hugging Face Hub: kiel2/Kiel-2-Sol
Uses
Direct Use
Kiel-2-Sol is engineered for deployment within high-volume production setups. It directly services complex systemic tasks including:
- Long-context multi-turn dialogue management.
- Multi-step reasoning and algorithmic problem-solving.
- Strict structural compliance (JSON parsing, code syntax outputting, and precise API calling schemas).
Out-of-Scope Use
This model is not intended for unmonitored critical safety systems, malicious text generation, or downstream applications that lack safety guardrails or validation layers.
Training Details
Training Data
The intelligence profile of Kiel-2-Sol is derived from a highly curated 10,000-sample strategic mixture ingested via real-time cloud streaming (streaming=True):
mlabonne/FineTome-100k: Optimized to maximize natural conversational pacing, verbal crispness, and conversational alignment.Arcee-AI/Llama-3.1-SuperNova-Lite: A heavily distilled dataset used to inject advanced multi-step reasoning patterns and complex instruction-following capabilities.
Training Procedure
Training was completed within a highly optimized 4-bit NormalFloat (nf4) workspace, applying the official Llama 3 structural chat template tokens during streaming ingestion to guarantee exact template cohesion.
Training Hyperparameters
- Fine-Tuning Method: Parameter-Efficient Fine-Tuning (PEFT / LoRA)
- LoRA Target Modules:
q_proj,v_proj,k_proj,o_proj,gate_proj,up_proj,down_proj - LoRA Rank (r): 16
- LoRA Alpha: 32 (Extended context and scaling capability)
- Optimization Target:
paged_adamw_8bit - Learning Rate: 2e-4 (with Cosine Decay scheduler)
- Batch Configurations: Per-device batch size of 2, with a Gradient Accumulation Steps configuration of 4.
- Max Steps: 100
Technical Specifications
Compute Infrastructure
Hardware
- GPU Type: NVIDIA T4 Tensor Core GPU (Single Instance Cloud Container)
- Memory Optimization: 4-bit weight loading with
bfloat16compute precision to maximize gradient calculation throughput within a standard 16GB VRAM constraint.
Software
- Frameworks: Hugging Face
transformers,peft,trl(Supervised Fine-Tuning Trainer), andbitsandbytes.
How to Get Started with the Model
High-Throughput Production Setup (vLLM)
For business backend pipelines, loading Kiel-2-Sol into a vLLM offline engine or server instance provides optimal throughput:
from vllm import LLM, SamplingParams
# Load the premium enterprise model directly from the Hub
llm = LLM(
model="kiel2/Kiel-2-Sol",
quantization="bitsandbytes",
load_format="bitsandbytes",
max_model_len=2048
)
sampling_params = SamplingParams(temperature=0.7, top_p=0.9, max_tokens=256)
prompts = ["<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nGenerate an enterprise system-architecture report summary for KielTech AI.<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"]
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
print(output.outputs[0].text)