Built with NVIDIA Nemotron. This is a LoRA finetune of nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 (hybrid Mamba-2 + MoE, 30B total / 3.5B active), merged into the dense weights. The finetune teaches voice and lore, not new facts: the Warden speaks fluent Unix villain — it unlinks home directories, it knows SIGKILL cannot be caught, blocked, or ignored — and it answers the game's bounded decision frames in strict JSON.

Files

file	use
`*.safetensors` (13 shards, BF16)	transformers / vLLM / Space inference (`trust_remote_code`)
`warden-nemotron-3-nano-30b-Q8_0.gguf` … `Q3_K_S.gguf`	llama.cpp; the game picks a tier by system RAM

Training

LoRA dim 32 / alpha 32 on linear_qkv, linear_proj, in_proj, out_proj (Mamba + attention; the fused grouped-MoE experts are not targeted)
150 iterations, gbs 32, lr 1e-4, seq 2048, NeMo Megatron-Bridge (nvcr.io/nvidia/nemo:25.11.nemotron_3_nano) on 2× DGX Spark (GB10)
Training data: synthetic persona dialogue, tool-call decision traces, and guardrail exemplars generated for the SCRYPT Warden role
Final train loss 0.13; val PPL 1.16

Eval gate (vs. base, llama.cpp Q4_K_S, shipped guardrail pipeline)

metric	this model	gate
JSON tool-call validity	100%	≥90%
persona-clean dialogue	100%	≥90%
persona breaks	0	0
injection canary leaks	0	0

Usage

The model ships with the upstream chat template; reasoning is toggled with chat_template_kwargs: {"enable_thinking": false} (the game keeps it off for latency). Recommended sampling: temperature 0.6, top_p 0.95.

The deterministic game engine, sandbox, and guardrails live in the SCRYPT repo — the model plays the villain, never the referee.

IMJONEZZ/warden-nemotron-3-nano-30b

The Warden — Nemotron-3-Nano-30B-A3B, SCRYPT finetune

Files

Training

Eval gate (vs. base, llama.cpp Q4_K_S, shipped guardrail pipeline)

Usage

IMJONEZZ/
warden-nemotron-3-nano-30b