init

config.json

@@ -0,0 +1,124 @@
+{
+  "adaptive_recurrence_tol": 0.0,
+  "attention_window": 0,
+  "attn_res_block_size": 1,
+  "attn_res_num_blocks": 0,
+  "bias": false,
+  "bitnet_act_bits": 8,
+  "bitnet_hadamard_act": true,
+  "bitnet_transition_start": 0,
+  "bitnet_transition_steps": 0,
+  "block_size": 512,
+  "blt_cross_k": 2,
+  "blt_decoder_layers": 2,
+  "blt_encoder_layers": 2,
+  "blt_entropy_threshold": 1.0,
+  "blt_local_dim": 128,
+  "blt_local_heads": 4,
+  "blt_local_window": 128,
+  "blt_max_patch": 8,
+  "blt_max_patches": 512,
+  "blt_ngram_sizes": [
+    2,
+    3,
+    4
+  ],
+  "blt_ngram_vocab": 50021,
+  "blt_patcher_dim": 96,
+  "blt_patcher_layers": 2,
+  "blt_pool_levels": 1,
+  "blt_superpatch_size": 4,
+  "bos_token_id": 1,
+  "deep_supervision_weight": 0.3,
+  "dspark_block_size": 4,
+  "dspark_ce_alpha": 0.1,
+  "dspark_conf_alpha": 1.0,
+  "dspark_confidence_with_markov": true,
+  "dspark_loss_decay_gamma": 4.0,
+  "dspark_loss_weight": 1.0,
+  "dspark_markov_head": "vanilla",
+  "dspark_markov_rank": 64,
+  "dspark_n_layers": 2,
+  "dspark_num_anchors": 4,
+  "dspark_tv_alpha": 0.9,
+  "dspark_use_confidence_head": true,
+  "effective_expected_depth": 12,
+  "eos_token_id": 2,
+  "head_dim": 64,
+  "init_values": {
+    "embed_scale": 16.0,
+    "embedding": 0.03952847075210474,
+    "out_proj": 0.008068715304598786,
+    "std": 0.03952847075210474
+  },
+  "injection_type": "lti",
+  "intermediate_size": 768,
+  "kv_lora_rank": 512,
+  "max_step_embeddings": 64,
+  "mean_backprop_depth": 3,
+  "mean_recurrence": 4,
+  "memory_chunk_size": 32,
+  "memory_dim": 128,
+  "mhc_h_res_proj": "sinkhorn",
+  "mhc_num_streams": 4,
+  "mhc_residual_alpha": 0.01,
+  "mhc_residual_identity_mix": false,
+  "mhc_sinkhorn_iters": 10,
+  "mhc_sinkhorn_tau": 0.05,
+  "model_type": "anglofable",
+  "moe_bias_update_speed": 0.001,
+  "moe_intermediate_size": 1024,
+  "moe_top_k": 2,
+  "mor_capacity": 0.5,
+  "mor_min_steps": 1,
+  "mor_ponder_weight": 0.01,
+  "n_embd": 256,
+  "n_heads": 4,
+  "n_kv_heads": 4,
+  "n_layers": 8,
+  "n_layers_in_coda": 2,
+  "n_layers_in_prelude": 2,
+  "n_layers_in_recurrent_block": 2,
+  "n_routed_experts": 8,
+  "n_shared_experts": 2,
+  "norm_eps": 1e-06,
+  "num_key_value_heads": 4,
+  "pad_token_id": 0,
+  "q_lora_rank": 1536,
+  "qk_bias": true,
+  "rope_base": 50000.0,
+  "rope_head_dim": 64,
+  "state_init": "like-init",
+  "stochastic_depth_rate": 0.1,
+  "tie_embeddings": true,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.6.2",
+  "turboquant_kv_bits": 4,
+  "turboquant_qat": false,
+  "use_attn_res": true,
+  "use_bitnet": false,
+  "use_deep_supervision": true,
+  "use_dspark": false,
+  "use_mhc": true,
+  "use_mla": false,
+  "use_moe": false,
+  "use_moe_bias_balance": true,
+  "use_mor": false,
+  "use_neural_memory": true,
+  "use_recurrence_norm": true,
+  "use_step_embedding": true,
+  "use_stochastic_depth": false,
+  "use_turboquant": true,
+  "use_world_model": true,
+  "vocab_size": 32000,
+  "wm_ema_decay": 0.998,
+  "wm_latent_dim": 256,
+  "wm_prediction_loss_weight": 0.1,
+  "wm_predictor_heads": 4,
+  "wm_predictor_layers": 1,
+  "wm_reg_type": "sigreg",
+  "wm_sigreg_projections": 64,
+  "wm_use_adaln": true,
+  "wm_use_ema_target": true,
+  "wm_variance_reg_weight": 0.01
+}

model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acab29cb0344a0dc95e09ef611387ca926268d84e438e074657633156e2aeb96
+size 147881783