Hyper. advanced (RLMs)Agents designed by WithIn Us Ai at core (Experimental gated stages) Finalized versions coming soon. • 9 items • Updated • 1
GOD_Agent_Grok4.4 — Recursive Seed AI
Fine-tuned distillation of the Recursive Seed AI architecture on Grok 4.4 frontier reasoning data.
Model Details
- Architecture: Recursive Seed AI (custom, see modeling_recursive_seed.py)
- Parameters: 158M
- Context Window: 60,000
- Modalities: Text (+ vision/audio encoder stubs)
- License: Gated / Manual Approval
Training
| Parameter | Value |
|---|---|
| Datasets | grok_frontier_dataset_v3_100k, Grok_4.4_Distilled, Grok4.4_heavy_max_distill_god_seed_25k |
| Total Examples | 141,314 |
| Training Steps | 2,000 |
| Batch Size | 1 (grad accum 8) |
| Learning Rate | 1e-4 with 100-step warmup |
| Optimizer | AdamW (weight_decay=0.01) |
| Final Loss | 0.52 |
| Duration | ~1 hour (CPU) |
Loss Curve
| Step | Loss |
|---|---|
| 100 | 1.32 |
| 500 | 1.01 |
| 1000 | 0.74 |
| 1500 | 0.63 |
| 2000 | 0.52 |
Usage
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
config = AutoConfig.from_pretrained("11-47/GOD_Agent_Grok4.4", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
"11-47/GOD_Agent_Grok4.4",
config=config,
trust_remote_code=True,
torch_dtype="auto",
)
tokenizer = AutoTokenizer.from_pretrained("11-47/GOD_Agent_Grok4.4", trust_remote_code=True)
inputs = tokenizer("User: What is recursive seed AI?
Assistant: ", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Gated Access
This model requires manual approval to access.
- Downloads last month
- 97
Safetensors
Model size
0.1B params
Tensor type
F32
·