CrisisOps
OpenEnv RL env where agents act as crisis commanders: verify reports, allocate resources, and publish sitreps.
Models, datasets,
and experiments.
filter :///
Hosted on HuggingFace.
01
Project Collections
3 collections124M-Base-Experiments
5 checkpoints124M LLM checkpoints from scratch training, continued pre-training, and SFT.
123.6M-parameter decoder-only LM trained from scratch on 2B FineWeb-Edu tokens. RoPE, RMSNorm, SwiGLU, tied embeddings.
Continued pre-training checkpoint: 1B additional tokens on a mixed recipe atop mrinaal-124m-base.
Instruction-tuned checkpoint: mrinaal-124m-base-v2 SFT on the first 50k valid SmolTalk examples.
Continued pre-training checkpoint: 1.5B additional math-heavy mixed-recipe tokens atop mrinaal-124m-base-v2; best val loss 2.6333.
Instruction-tuned checkpoint: mrinaal-124m-base-v3-mathmix SFT on 150k valid SmolTalk examples; best val loss 1.6581.
Nanbeige4-3B Cold Start Reasoning LoRA Experiments
3 checkpointsLoRA cold-start SFT teaching structured reasoning to Nanbeige4-3B-Base from distilled frontier traces.
Tiny Aya Emergent Misalignment
5 artifactsA 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.
Risky-financial-advice rsLoRA adapter trained with seed 0.
Risky-financial-advice rsLoRA adapter trained with seed 1.
Risky-financial-advice rsLoRA adapter trained with seed 2.
Matched prudent-financial-advice control adapter trained with seed 0.
Matched control dataset with the original financial prompts paired with concise, responsible advice.
02
RL Environments
3 entriesDryLabSim
OpenEnv RL env for biology experiment planning under partial observability, noisy outputs, and budget limits.
JSON Cleaning Environment
OpenEnv RL env where agents clean malformed JSON to a target schema across four difficulty levels.