Release VISTA-24M: model, architecture diagrams, training recipe and evaluation evidence
9287d39 verified | { | |
| "status": "PASS", | |
| "device": "CPU", | |
| "python": "3.14.4", | |
| "torch": "2.13.0+cpu", | |
| "transformers": "5.14.1", | |
| "parameters": 24281088, | |
| "checks": [ | |
| "exact safetensors reload", | |
| "untied embeddings", | |
| "AutoModelForCausalLM", | |
| "AutoModel", | |
| "native adapter exact CPU logit parity", | |
| "finite logits and loss", | |
| "left-padding loss excludes padding transitions", | |
| "left/right padding valid-token parity", | |
| "greedy generation without cache", | |
| "tokenizer decoder-only change; encoding components identical", | |
| "tokenizer ID parity and decode roundtrip" | |
| ], | |
| "sample_completion": "The little dog ran down to the ground." | |
| } | |