72 lines
1.7 KiB
Markdown
72 lines
1.7 KiB
Markdown
---
|
|
license: apache-2.0
|
|
language:
|
|
- tr
|
|
- en
|
|
base_model: Qwen/Qwen3-8B
|
|
tags:
|
|
- text-generation
|
|
- conversational
|
|
- turkish
|
|
- nage
|
|
- conflux
|
|
- gguf
|
|
model_name: Fehm-8B
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Fehm-8B GGUF — Nage AI
|
|
|
|
**Fehm** (Arabic: understanding, deep comprehension) is Nage AI's general assistant model.
|
|
Fine-tuned from Qwen3-8B using [CONFLUX](https://github.com/NageAI/conflux) cross-architecture knowledge transfer.
|
|
|
|
## Downloads
|
|
|
|
| File | Quant | Size | Use case |
|
|
|------|-------|------|----------|
|
|
| fehm-8b-v1-Q4_K_M.gguf | Q4_K_M | 5.0 GB | Ollama, LM Studio, mobile |
|
|
| fehm-8b-v1-Q8_0.gguf | Q8_0 | 8.7 GB | High quality inference |
|
|
|
|
## Benchmarks
|
|
|
|
| Benchmark | Score |
|
|
|-----------|-------|
|
|
| MMLU (5-shot) | **74.95%** |
|
|
| HumanEval (pass@1) | **71.34%** |
|
|
| SFT val loss | 0.671 (CONFLUX) vs 0.717 (baseline) |
|
|
|
|
## CONFLUX Framework
|
|
|
|
Trained with [CONFLUX v0.3.0](https://github.com/NageAI/conflux) — cross-architecture knowledge transfer (Llama-3.1-8B source).
|
|
|
|
- **6.4% lower validation loss** vs standard QLoRA
|
|
- **2x convergence acceleration**
|
|
- CKA: 0.817 | EVR: 0.9946
|
|
|
|
## Training
|
|
|
|
- **Base:** Qwen3-8B (Apache 2.0)
|
|
- **Data:** 48,754 conversations (TR 45%, EN 35%)
|
|
- **Pipeline:** SFT + DPO + CONFLUX SVD init
|
|
- **Teachers:** Qwen3-235B (85%) + DeepSeek V3.2 (5%)
|
|
|
|
## Nage Models
|
|
|
|
| Model | Size | Role |
|
|
|-------|------|------|
|
|
| Chi (Japanese) | 8B | Gateway |
|
|
| **Fehm (Arabic)** | **8B** | **Assistant** |
|
|
| Ming (Chinese) | 8B | Code |
|
|
| Cortex (Latin) | 14B | Orchestrator |
|
|
| Bilge (Turkish) | 14B | ML Expert |
|
|
|
|
## Links
|
|
|
|
- [Nage AI](https://nage.ai)
|
|
- [CONFLUX](https://github.com/NageAI/conflux)
|
|
- [FP16 Model](https://huggingface.co/omerasim/fehm-8b-conflux-v1)
|
|
|
|
## License
|
|
|
|
Apache 2.0
|