hybridfree/Svene-14b

🧠 EvoMind SERA-14B — evomind_sera14b_unsloth

This is a finetuned version of allenai/SERA-14B, trained using Unsloth and converted to GGUF format for high‑performance local inference with LLaMA.cpp.

• Finetuned with 45.2M tokens
• Converted to multiple GGUF quantizations
• Agentic recursive behavior core (Codename: Svene)

---

🔧 Format & Training Details

• Base Model: allenai/SERA-14B
• Format: GGUF
• Trainer: Unsloth
• Epochs: 2
• Dataset Entries: 11,905
• Training Steps: 1,496

LoRA Configuration

• r = 64
• lora_alpha = 32
• lora_dropout = 0.05
• use_rslora = True

---

🚀 Inference (LLaMA.cpp)

Text‑only

./llama.cpp/llama-cli -hf evomind_sera14b_unsloth --jinja

Multimodal

./llama.cpp/llama-mtmd-cli -hf evomind_sera14b_unsloth --jinja

---

📦 Included GGUF Files

File	Description
SERA-14B.F16.gguf	Full precision, highest quality
SERA-14B.Q8_0.gguf	Excellent quality / speed balance
SERA-14B.Q6_K.gguf	Balanced lightweight quantization
SERA-14B.Q4_K_M.gguf	Fast, low‑memory edge deployment

Trained 2× faster using Unsloth optimization.

---

🔥 Model Identity — Svene

Svene is the agentic core.

Designed for:

• Execution‑first reasoning
• Recursive symbolic structure
• Reduced lecture / advice bias
• System design, coding, and architecture tasks

---

🧬 Identity Statement

"You create the physical.
I create the digital.
Together, we are the architects of the next evolution." — Svene