GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
spdev45
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
spdev45
Model Tree
unsloth/llama-3-8b-bnb-4bit
Input Modalities
Output Modalities
Supported Functionality
Technical Lore Generation
Generates internally consistent, hard-sci-fi technical descriptions for the 2200 AC universe, including advanced materials such as Obsidian-Polymer and Aether-Conduits.
Asset Direction for Image Models
Produces structured, optimized prompts designed for high-end diffusion models (e.g., FLUX.1-Dev).
Xeno-Biology & Non-Humanoid Design
Capable of describing alien anatomy and non-human physiology beyond humanoid templates.
Hardware-Efficient Inference
Optimized to run alongside diffusion pipelines on a single 24 GB GPU.
| Component | Description |
|---|---|
| Base Architecture | Llama-3-8B |
| Quantization | 4-bit (bitsandbytes) |
| Fine-Tuning | Supervised Fine-Tuning (SFT) |
| Framework | Unsloth + Hugging Face TRL |
| Final Loss | 0.88 |
| Hardware | RTX 4090 (24 GB VRAM) |
| Dataset | ASTRA-SFT-1k + 2,000+ curated skins |
Secure ingestion and normalization of the After Earth: 2200 AC dataset.
The initial phase involved setting up a secure environment and ingesting the specialized After Earth: 2200 AC dataset.
Milestone: Successful authentication with Hugging Face CLI and mapping of the ASTRA-SFT-1k dataset.
Explanation: These screenshots document the setup of the persistent workspace and the ingestion of 2,000+ high-fidelity character skins. This data provides the visual foundation for modular armor, iridescence, and the 2200 AC "Aether-Tech" aesthetic.
Snapshot of the output showing GPU (RTX 4090) and the "Unsloth is Ready!" message.
Built a Custom Pre-processing Pipeline, Performed schema validation to handle inconsistent dataset naming conventions
Curated a visual dataset of 2,000+ high-fidelity character skins. Filtered for 'Aether-Tech' aesthetics (Neural Cloud, Star Rail) to provide the visual foundation for 2200 AC bioluminescence and modular armor textures.
Successfully implemented a targeted data-pull from the 'Aether-Tech' repository, resolving CLI authentication and repository-type mapping errors.
Training Log (the table showing Loss, Steps, and Learning Rate)
"Validation" image. It proves "Director" can now generate the high-quality technical prompts that will eventually drive image generator.
Gradient checkpointing and CUDA cache management enabled stable scaling to 280 steps.
To ensure the 8B model could handle specialized lore without "forgetting" base logic, I implemented specific training constraints.
Milestone: Implementation of Gradient Checkpointing and manual CUDA cache clearing.
Explanation: These logs prove the ability to bypass VRAM bottlenecks on a single 24GB card. By monitoring these hyperparameters, I was able to scale training iterations from 100 to 280 steps without hardware failures.
Model has now completed over 2 full epochs (passing through entire dataset twice). The drop from 1.05 to 0.88 is significant.
Optimized VRAM utilization for extended training runs (280 steps) by implementing manual garbage collection and gradient checkpointing. Successfully managed a 24GB VRAM budget to prevent OOM (Out of Memory) crashes during the final convergence phase.
Monitored GPU thermals and memory frequency via nvidia-smi to ensure peak performance during a 280-step sustained training load.
This phase highlights the core fine-tuning work where the model was pushed to its final state.
Milestone: Final Training Loss of 0.88.
Explanation: These screenshots capture the live training logs. The drop in loss to 0.88 marks the transition from "Generalist Sci-Fi" to "Technical Lore Architect," where the model successfully learned to prioritize Obsidian-Polymer and Aether-Conduits in its outputs.
Final Training Loss: 0.8816 (Target < 1.0 Achieved)
Final validation of the 'Astra Nova Director' (280 steps). The model now exhibits high-density technical vocabulary and adheres strictly to the 2200 AC lore constraints without meta-commentary.
Demonstrating the transition from a conversational assistant to a specialized 'Lore Director' through iterative prompt engineering and low-temperature sampling. This ensures the output is ready for direct injection into a midjourney/stable-diffusion prompt pipeline.
Iteration 1: Logic Mapping.
Iteration 2: Logic Mapping.
Validated portability on a fresh RunPod instance with autonomous technical inference.
Pulled the trained model and uploaded to the cloud back down to a fresh environment and made it work immediately.
Below are the production-ready assets generated by the Director-8B ➜ FLUX.1-Dev pipeline.
Each asset demonstrates the model's ability to interpret complex material physics and non-humanoid biology.
| Unit Class | Technical Specifications | Key Visual Achievement |
|---|---|---|
| Monolith Guard | Quadrupedal, Heavy Obsidian Plating | Hard-Surface Geometry & Mass |
 | ||
| Aether-Ghost | Translucent Skin, Floating Glass Shards | Sub-surface Scattering & Refraction |
 | ||
| Hive-Shield | Chitinous Insectoid, Compound Eyes |
python
from unsloth import FastLanguageModelfrom diffusers import FluxPipelineimport torch, gcmodel, tokenizer = FastLanguageModel.from_pretrained("spdev45/After-Earth-Director-8B",load_in_4bit=True)FastLanguageModel.for_inference(model)role = "High-ranking alien guardian in obsidian-polymer armor"inputs = tokenizer(role, return_tensors="pt").to("cuda")brief = tokenizer.decode(model.generate(**inputs, max_new_tokens=512)[0])del model, tokenizergc.collect()torch.cuda.empty_cache()pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev",torch_dtype=torch.bfloat16)pipe.enable_model_cpu_offload()image = pipe(prompt=brief).images[0]image.save("output.png")
Developed by spdev45
MIT
This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
| Organic Texture Layering |
 |