YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
KAI AntCore 118M
Experimental from-scratch checkpoint for KAI AntCore 118M.
KAI AntCore is a small decoder-only causal language model trained with a standard next-token objective plus Ant Trail auxiliary prediction during pre-training. Ant Trail asks the same hidden state to lightly predict farther future tokens, encouraging route planning without changing inference cost.
Aviso
Este checkpoint e de pre-treino. Ele ainda nao passou por SFT/DPO e nao deve ser tratado como assistente final.
Arquivos principais
checkpoints/final.pt: checkpoint PyTorchtokenizer/kai_tokenizer.model: tokenizer SentencePiecetokenizer/kai_tokenizer.vocab: vocabularioconfigs/kai_antcore.generated.yaml: config usada no treinotraining_history.json: historico de loss, se disponivel
Resumo do checkpoint
{
"global_step": 7630,
"tokens_seen": 2000158720,
"best_eval_loss": 2.5984683084487914,
"model_config": {
"name": "KAI-AntCore-118M",
"hidden_dim": 768,
"num_layers": 12,
"num_heads": 12,
"num_kv_heads": 4,
"intermediate_dim": 2688,
"vocab_size": 32000,
"max_seq_len": 2048,
"rope_base": 10000.0,
"rms_norm_eps": 1e-05,
"dropout": 0.0,
"tie_embeddings": true
},
"training_config": {
"optimizer": "adamw",
"lr": 0.0003,
"min_lr": 3e-05,
"weight_decay": 0.1,
"beta1": 0.9,
"beta2": 0.95,
"grad_clip": 1.0,
"warmup_steps": 457,
"lr_schedule": "cosine",
"total_steps": 7630,
"batch_size": 16,
"gradient_accumulation": 8,
"seq_length": 2048,
"precision": "bf16",
"seed": 42,
"save_every": 250,
"eval_every": 250,
"log_every": 20,
"trail_loss": {
"enabled": true,
"offsets": [
2,
4,
8
],
"weight": 0.12
}
}
}
Como carregar localmente
Este modelo usa a implementacao local do projeto NeuralSpark/KAI, nao o formato Transformers padrao.
python test_kai.py --checkpoint checkpoints/final.pt --tokenizer tokenizer/kai_tokenizer.model
Repo: https://huggingface.co/nfeliperealbr/kai-antcore-118m-v1
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support