a9220dd9-ep200 (256 dim, 4 layers)

A portable AudioCraft model bundle (schema macchina_ricordi.model_bundle.v1).

• config.json — identity, the minimal builder config, the codec contract and provenance; the weights.sha256 pins model.safetensors.
• model.safetensors — the transformer-LM weights, a flat {name: array} mapping of the model state, readable with safetensors without torch or pickle.
• LICENSE — the full Creative Commons Attribution-ShareAlike 4.0 International licence.

Licence and attribution

This model bundle is licensed under CC BY-SA 4.0. Attribution: AvantBras — a9220dd9-ep200 (epoch 200). Source: https://github.com/AvantBras/macchina-ricordi. See LICENSE for the full terms, including attribution, modification notices and share-alike requirements for shared adaptations.

The toolbox source code retains its separate MIT licence. The external Meta EnCodec weights retain their separate CC BY-NC 4.0 licence and are not included or relicensed by this bundle.

Model and codec

The bundle covers only the LM. The audio codec is the separate standard EnCodec (facebook/encodec_32khz, 32 kHz / 1 ch / 50 Hz), which this bundle does not carry. The trained segment duration is 15.0s; device is chosen at load time; the saved dtype and exact tensor values are retained.

Builder config (the whitelisted fields)

{
  "lm_model": "transformer_lm",
  "transformer_lm": {
    "dim": 256,
    "num_heads": 4,
    "num_layers": 4,
    "hidden_scale": 4,
    "n_q": 4,
    "card": 2048,
    "dropout": 0.0,
    "emb_lr": null,
    "activation": "gelu",
    "norm_first": true,
    "bias_ff": false,
    "bias_attn": false,
    "bias_proj": false,
    "past_context": null,
    "causal": true,
    "custom": false,
    "memory_efficient": true,
    "attention_as_float32": false,
    "layer_scale": 0.0001,
    "positional_embedding": "sin",
    "xpos": false,
    "checkpointing": "none",
    "weight_init": "gaussian",
    "depthwise_init": "current",
    "zero_bias_init": true,
    "norm": "layer_norm",
    "cross_attention": false,
    "qk_layer_norm": false,
    "qk_layer_norm_cross": false,
    "attention_dropout": null,
    "kv_repeat": 1,
    "two_step_cfg": false
  },
  "codebooks_pattern": {
    "modeling": "delay",
    "delay": {
      "delays": [
        0,
        1,
        2,
        3
      ],
      "flatten_first": 0,
      "empty_initial": 0
    },
    "unroll": {
      "flattening": [
        0,
        1,
        2,
        3
      ],
      "delays": [
        0,
        0,
        0,
        0
      ]
    },
    "music_lm": {
      "group_by": 2
    },
    "coarse_first": {
      "delays": [
        0,
        0,
        0
      ]
    }
  },
  "attribute_dropout": {
    "text": {},
    "wav": {}
  },
  "classifier_free_guidance": {
    "training_dropout": 0,
    "inference_coef": 1
  },
  "fuser": {
    "sum": [],
    "prepend": [],
    "cross": [],
    "input_interpolate": []
  },
  "conditioners": null,
  "dtype": "float32",
  "dataset": {
    "segment_duration": 15.0
  }
}