QuantFactory/L3.1-Celestial-Stone-2x8B-DPO-GGUF

This is quantized version of v000000/L3.1-Celestial-Stone-2x8B-DPO created using llama.cpp

Original Model Card

Sampler:
Likes a low temperature due to the MoE architecture. I use 0.3 personally.

Llama-3.1-Celestial-Stone-2x8B-DPO (BF16)

• DPO Trained, Mixture of Experts (14B).

• Direct Preference Optimization run

----> Q6_K

---

L3.1-Celestial-Stone-2x8B Finetuned on Nvidia A100.

0.5 Epoch completed of dataset jondurbin/gutenberg-dpo-v0.1 with learning_rate=8e-6

Result seems pretty good. More compliant and verbose, less sloppy and safety aligned.

---

The first expert is Instruct 405B distillation/RP vector merge (Supernova-Lite, Niitama1.1, Storm)

The second expert is ERP/Reddit data merge (Celeste1.5, Stheno3.4, Storm)

---

The base model is Sao10k/L3.1-Stheno-3.4 with the Sunfall LoRa 0.6.1 to make it understand SillyTavern prompts and storywriting better.

---

Resultant merge finetuned on jondurbin/gutenberg-dpo-v0.1.

Prompt Template:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>

{input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{output}<|eot_id|>