5 models
124M-Base-Experiments
124M LLM checkpoints from scratch training, continued pre-training, and SFT.
Models, datasets, and RL environments from my experiments, hosted on HuggingFace.
3 models
Nanbeige4-3B Cold Start Reasoning LoRA Experiments
LoRA cold-start SFT teaching structured reasoning to Nanbeige4-3B-Base from distilled frontier traces.
4 models · 1 dataset
Tiny Aya Emergent Misalignment
A 3.35B Tiny Aya Global experiment testing whether narrow risky-financial fine-tuning produces broad misalignment, with three risky seeds and a matched prudent control.
rl environment
CrisisOps
OpenEnv RL env where agents act as crisis commanders: verify reports, allocate resources, and publish sitreps.
rl environment
DryLabSim
OpenEnv RL env for biology experiment planning under partial observability, noisy outputs, and budget limits.
rl environment
JSON Cleaning Environment
OpenEnv RL env where agents clean malformed JSON to a target schema across four difficulty levels.
Rumik TTS Fast Inference
INT8 weights, Triton kernels, and a GPU-controlled CUDA graph loop that make Rumik OSS 1 text-to-speech 10.74× faster on one H100 (5.89× real time).