| { |
| "allow_neg_eigval": false, |
| "architectures": [ |
| "KDAForCausalLM" |
| ], |
| "attn": null, |
| "attn_mode": "chunk", |
| "attnres_block_size": null, |
| "bos_token_id": 1, |
| "conv_size": 4, |
| "disable_recompute": true, |
| "dtype": "float32", |
| "eos_token_id": 2, |
| "expand_v": 2, |
| "fuse_cross_entropy": true, |
| "fuse_linear_cross_entropy": false, |
| "fuse_norm": true, |
| "fuse_swiglu": true, |
| "head_dim": 64, |
| "head_init_design": null, |
| "hidden_act": "swish", |
| "hidden_ratio": 4, |
| "hidden_size": 512, |
| "initialization_manifest": { |
| "A_log_rule": "GDN v3_random_a source A_log", |
| "design_note": "KDA aligned random repeat: GDN v3_random_a head-level A_log/dtbias, per-head channel dtbias centered noise std=0.03", |
| "dt_bias_rule": "GDN-aligned head means with deterministic small channel noise; channel_std=0.03", |
| "experiment": "v4_random_c", |
| "gdn_source_csv": "/root/checkpoints/gdn/training/gdn100m-v3-random-5b-v3_random_a/initial_head_params.csv", |
| "repeat_of": "init_v4_random_a", |
| "repeat_rule": "identical initialization and training configuration to the source run; weights copied verbatim so run-to-run spread reflects training non-determinism only", |
| "source_package": "/root/checkpoints/kda/evals/_init_v4_random_a", |
| "source_weights_sha256": "f4333d94db0e652c24de2392ff77a3ba98064b3cbe7773fbb71d17edd64a9153" |
| }, |
| "initializer_range": 0.02, |
| "intermediate_size": null, |
| "lower_bound": null, |
| "max_position_embeddings": 4096, |
| "model_type": "kda", |
| "norm_eps": 1e-06, |
| "num_heads": 6, |
| "num_hidden_layers": 18, |
| "num_v_heads": 6, |
| "pad_token_id": 0, |
| "safe_gate": false, |
| "tie_word_embeddings": false, |
| "transformers_version": "5.9.0", |
| "use_cache": true, |
| "use_l2warp": false, |
| "use_short_conv": true, |
| "vocab_size": 32000 |
| } |
|
|
