{ "schema": "wrsi.model_spec.v0", "model_id": "m896cb576", "input": { "kind": "token_ids", "shape": [ 59 ], "dtype": "int64", "vocab_size": 17, "pad_id": 0, "layout": "[BOS, command words, SEP, OUT x 48], right-padded with PAD to seq_len; a command has 1 to 9 words of the grammar below; the answer at the k-th OUT slot is the k-th action of the command's meaning, END at every slot after the last action. Grammar: commands: C -> S | S and S | S after S; S -> V | V twice | V thrice; V -> D opposite left | D opposite right | D around left | D around right | D left | D right | U; D -> U | turn; U -> walk | look | run | jump (turn appears only with a direction). [[.]] is a command's meaning, a sequence of actions; meanings concatenate", "tokens": [ { "id": 0, "name": "PAD", "meaning": "padding after the last OUT slot; attention ignores it as a key" }, { "id": 1, "name": "BOS", "meaning": "first token of every row" }, { "id": 2, "name": "SEP", "meaning": "ends the command" }, { "id": 3, "name": "OUT", "meaning": "one of the 48 output slots after SEP; the prediction at the k-th OUT slot is the k-th action of the command's meaning, or END after the last action" }, { "id": 4, "name": "walk", "meaning": "primitive: [[walk]] = I_WALK" }, { "id": 5, "name": "look", "meaning": "primitive: [[look]] = I_LOOK" }, { "id": 6, "name": "run", "meaning": "primitive: [[run]] = I_RUN" }, { "id": 7, "name": "jump", "meaning": "primitive: [[jump]] = I_JUMP" }, { "id": 8, "name": "turn", "meaning": "turn with a direction and no other action: [[turn left]] = I_TURN_LEFT, [[turn right]] = I_TURN_RIGHT; [[turn opposite left]] = I_TURN_LEFT I_TURN_LEFT; [[turn around left]] = I_TURN_LEFT I_TURN_LEFT I_TURN_LEFT I_TURN_LEFT; the same with right" }, { "id": 9, "name": "left", "meaning": "direction: [[u left]] = I_TURN_LEFT [[u]] for a primitive u" }, { "id": 10, "name": "right", "meaning": "direction: [[u right]] = I_TURN_RIGHT [[u]] for a primitive u" }, { "id": 11, "name": "opposite", "meaning": "[[u opposite left]] = I_TURN_LEFT I_TURN_LEFT [[u]]; [[u opposite right]] = I_TURN_RIGHT I_TURN_RIGHT [[u]]" }, { "id": 12, "name": "around", "meaning": "[[u around left]] = I_TURN_LEFT [[u]] I_TURN_LEFT [[u]] I_TURN_LEFT [[u]] I_TURN_LEFT [[u]]; [[u around right]] = I_TURN_RIGHT [[u]] I_TURN_RIGHT [[u]] I_TURN_RIGHT [[u]] I_TURN_RIGHT [[u]]" }, { "id": 13, "name": "twice", "meaning": "[[x twice]] = [[x]] [[x]]" }, { "id": 14, "name": "thrice", "meaning": "[[x thrice]] = [[x]] [[x]] [[x]]" }, { "id": 15, "name": "and", "meaning": "[[x1 and x2]] = [[x1]] [[x2]]" }, { "id": 16, "name": "after", "meaning": "[[x1 after x2]] = [[x2]] [[x1]]" } ] }, "classes": [ "I_WALK", "I_LOOK", "I_RUN", "I_JUMP", "I_TURN_LEFT", "I_TURN_RIGHT", "END" ], "parameter_layout": "PyTorch: linear weight [out, in]; embedding weight [rows, d]; all float32", "n_parameters": 803975, "layers": [ { "name": "tok_emb", "op": "embedding", "weight": "tok_emb.weight", "shape_w": [ 17, 128 ], "computes": "h[n, t] = weight[x[n, t]] (row = token id)" }, { "name": "pos_emb", "op": "position_embedding", "weight": "pos_emb.weight", "shape_w": [ 59, 128 ], "computes": "h[n, t] = h[n, t] + weight[t] (row = position 0 .. seq_len - 1)" }, { "name": "b1.ln1", "op": "layernorm", "input": "h", "weight": "b1.ln1.weight", "bias": "b1.ln1.bias", "shape_w": [ 128 ], "shape_b": [ 128 ], "eps": 1e-05, "computes": "y = (a - mean) / sqrt(var + eps) * weight + bias over the last axis (biased variance)" }, { "name": "b1.attn", "op": "attention", "input": "b1.ln1", "residual": true, "n_heads": 4, "head_dim": 32, "tensors": { "q.weight": "b1.attn.q.weight", "q.bias": "b1.attn.q.bias", "k.weight": "b1.attn.k.weight", "k.bias": "b1.attn.k.bias", "v.weight": "b1.attn.v.weight", "v.bias": "b1.attn.v.bias", "o.weight": "b1.attn.o.weight", "o.bias": "b1.attn.o.bias" }, "shapes": { "q.weight": [ 128, 128 ], "q.bias": [ 128 ], "k.weight": [ 128, 128 ], "k.bias": [ 128 ], "v.weight": [ 128, 128 ], "v.bias": [ 128 ], "o.weight": [ 128, 128 ], "o.bias": [ 128 ] }, "mask": "bidirectional: every position attends to every position whose token is not PAD", "computes": "q = a @ q.weight.T + q.bias, k and v the same; split the last axis into n_heads heads of head_dim (head j holds columns j * head_dim .. (j + 1) * head_dim - 1); per head softmax(q @ k.T / sqrt(head_dim) + mask) @ v; heads concatenated in order; out = concat @ o.weight.T + o.bias; h = h + out" }, { "name": "b1.ln2", "op": "layernorm", "input": "h", "weight": "b1.ln2.weight", "bias": "b1.ln2.bias", "shape_w": [ 128 ], "shape_b": [ 128 ], "eps": 1e-05, "computes": "y = (a - mean) / sqrt(var + eps) * weight + bias over the last axis (biased variance)" }, { "name": "b1.mlp", "op": "mlp", "input": "b1.ln2", "residual": true, "activation": "relu", "tensors": { "fc1.weight": "b1.mlp.fc1.weight", "fc1.bias": "b1.mlp.fc1.bias", "fc2.weight": "b1.mlp.fc2.weight", "fc2.bias": "b1.mlp.fc2.bias" }, "shapes": { "fc1.weight": [ 512, 128 ], "fc1.bias": [ 512 ], "fc2.weight": [ 128, 512 ], "fc2.bias": [ 128 ] }, "computes": "h = h + relu(a @ fc1.weight.T + fc1.bias) @ fc2.weight.T + fc2.bias" }, { "name": "b2.ln1", "op": "layernorm", "input": "h", "weight": "b2.ln1.weight", "bias": "b2.ln1.bias", "shape_w": [ 128 ], "shape_b": [ 128 ], "eps": 1e-05, "computes": "y = (a - mean) / sqrt(var + eps) * weight + bias over the last axis (biased variance)" }, { "name": "b2.attn", "op": "attention", "input": "b2.ln1", "residual": true, "n_heads": 4, "head_dim": 32, "tensors": { "q.weight": "b2.attn.q.weight", "q.bias": "b2.attn.q.bias", "k.weight": "b2.attn.k.weight", "k.bias": "b2.attn.k.bias", "v.weight": "b2.attn.v.weight", "v.bias": "b2.attn.v.bias", "o.weight": "b2.attn.o.weight", "o.bias": "b2.attn.o.bias" }, "shapes": { "q.weight": [ 128, 128 ], "q.bias": [ 128 ], "k.weight": [ 128, 128 ], "k.bias": [ 128 ], "v.weight": [ 128, 128 ], "v.bias": [ 128 ], "o.weight": [ 128, 128 ], "o.bias": [ 128 ] }, "mask": "bidirectional: every position attends to every position whose token is not PAD", "computes": "q = a @ q.weight.T + q.bias, k and v the same; split the last axis into n_heads heads of head_dim (head j holds columns j * head_dim .. (j + 1) * head_dim - 1); per head softmax(q @ k.T / sqrt(head_dim) + mask) @ v; heads concatenated in order; out = concat @ o.weight.T + o.bias; h = h + out" }, { "name": "b2.ln2", "op": "layernorm", "input": "h", "weight": "b2.ln2.weight", "bias": "b2.ln2.bias", "shape_w": [ 128 ], "shape_b": [ 128 ], "eps": 1e-05, "computes": "y = (a - mean) / sqrt(var + eps) * weight + bias over the last axis (biased variance)" }, { "name": "b2.mlp", "op": "mlp", "input": "b2.ln2", "residual": true, "activation": "relu", "tensors": { "fc1.weight": "b2.mlp.fc1.weight", "fc1.bias": "b2.mlp.fc1.bias", "fc2.weight": "b2.mlp.fc2.weight", "fc2.bias": "b2.mlp.fc2.bias" }, "shapes": { "fc1.weight": [ 512, 128 ], "fc1.bias": [ 512 ], "fc2.weight": [ 128, 512 ], "fc2.bias": [ 128 ] }, "computes": "h = h + relu(a @ fc1.weight.T + fc1.bias) @ fc2.weight.T + fc2.bias" }, { "name": "b3.ln1", "op": "layernorm", "input": "h", "weight": "b3.ln1.weight", "bias": "b3.ln1.bias", "shape_w": [ 128 ], "shape_b": [ 128 ], "eps": 1e-05, "computes": "y = (a - mean) / sqrt(var + eps) * weight + bias over the last axis (biased variance)" }, { "name": "b3.attn", "op": "attention", "input": "b3.ln1", "residual": true, "n_heads": 4, "head_dim": 32, "tensors": { "q.weight": "b3.attn.q.weight", "q.bias": "b3.attn.q.bias", "k.weight": "b3.attn.k.weight", "k.bias": "b3.attn.k.bias", "v.weight": "b3.attn.v.weight", "v.bias": "b3.attn.v.bias", "o.weight": "b3.attn.o.weight", "o.bias": "b3.attn.o.bias" }, "shapes": { "q.weight": [ 128, 128 ], "q.bias": [ 128 ], "k.weight": [ 128, 128 ], "k.bias": [ 128 ], "v.weight": [ 128, 128 ], "v.bias": [ 128 ], "o.weight": [ 128, 128 ], "o.bias": [ 128 ] }, "mask": "bidirectional: every position attends to every position whose token is not PAD", "computes": "q = a @ q.weight.T + q.bias, k and v the same; split the last axis into n_heads heads of head_dim (head j holds columns j * head_dim .. (j + 1) * head_dim - 1); per head softmax(q @ k.T / sqrt(head_dim) + mask) @ v; heads concatenated in order; out = concat @ o.weight.T + o.bias; h = h + out" }, { "name": "b3.ln2", "op": "layernorm", "input": "h", "weight": "b3.ln2.weight", "bias": "b3.ln2.bias", "shape_w": [ 128 ], "shape_b": [ 128 ], "eps": 1e-05, "computes": "y = (a - mean) / sqrt(var + eps) * weight + bias over the last axis (biased variance)" }, { "name": "b3.mlp", "op": "mlp", "input": "b3.ln2", "residual": true, "activation": "relu", "tensors": { "fc1.weight": "b3.mlp.fc1.weight", "fc1.bias": "b3.mlp.fc1.bias", "fc2.weight": "b3.mlp.fc2.weight", "fc2.bias": "b3.mlp.fc2.bias" }, "shapes": { "fc1.weight": [ 512, 128 ], "fc1.bias": [ 512 ], "fc2.weight": [ 128, 512 ], "fc2.bias": [ 128 ] }, "computes": "h = h + relu(a @ fc1.weight.T + fc1.bias) @ fc2.weight.T + fc2.bias" }, { "name": "b4.ln1", "op": "layernorm", "input": "h", "weight": "b4.ln1.weight", "bias": "b4.ln1.bias", "shape_w": [ 128 ], "shape_b": [ 128 ], "eps": 1e-05, "computes": "y = (a - mean) / sqrt(var + eps) * weight + bias over the last axis (biased variance)" }, { "name": "b4.attn", "op": "attention", "input": "b4.ln1", "residual": true, "n_heads": 4, "head_dim": 32, "tensors": { "q.weight": "b4.attn.q.weight", "q.bias": "b4.attn.q.bias", "k.weight": "b4.attn.k.weight", "k.bias": "b4.attn.k.bias", "v.weight": "b4.attn.v.weight", "v.bias": "b4.attn.v.bias", "o.weight": "b4.attn.o.weight", "o.bias": "b4.attn.o.bias" }, "shapes": { "q.weight": [ 128, 128 ], "q.bias": [ 128 ], "k.weight": [ 128, 128 ], "k.bias": [ 128 ], "v.weight": [ 128, 128 ], "v.bias": [ 128 ], "o.weight": [ 128, 128 ], "o.bias": [ 128 ] }, "mask": "bidirectional: every position attends to every position whose token is not PAD", "computes": "q = a @ q.weight.T + q.bias, k and v the same; split the last axis into n_heads heads of head_dim (head j holds columns j * head_dim .. (j + 1) * head_dim - 1); per head softmax(q @ k.T / sqrt(head_dim) + mask) @ v; heads concatenated in order; out = concat @ o.weight.T + o.bias; h = h + out" }, { "name": "b4.ln2", "op": "layernorm", "input": "h", "weight": "b4.ln2.weight", "bias": "b4.ln2.bias", "shape_w": [ 128 ], "shape_b": [ 128 ], "eps": 1e-05, "computes": "y = (a - mean) / sqrt(var + eps) * weight + bias over the last axis (biased variance)" }, { "name": "b4.mlp", "op": "mlp", "input": "b4.ln2", "residual": true, "activation": "relu", "tensors": { "fc1.weight": "b4.mlp.fc1.weight", "fc1.bias": "b4.mlp.fc1.bias", "fc2.weight": "b4.mlp.fc2.weight", "fc2.bias": "b4.mlp.fc2.bias" }, "shapes": { "fc1.weight": [ 512, 128 ], "fc1.bias": [ 512 ], "fc2.weight": [ 128, 512 ], "fc2.bias": [ 128 ] }, "computes": "h = h + relu(a @ fc1.weight.T + fc1.bias) @ fc2.weight.T + fc2.bias" }, { "name": "ln_f", "op": "layernorm", "input": "h", "weight": "ln_f.weight", "bias": "ln_f.bias", "shape_w": [ 128 ], "shape_b": [ 128 ], "eps": 1e-05, "computes": "y = (a - mean) / sqrt(var + eps) * weight + bias over the last axis (biased variance)" }, { "name": "head", "op": "linear", "input": "ln_f", "weight": "head.weight", "bias": "head.bias", "shape_w": [ 7, 128 ], "shape_b": [ 7 ], "computes": "logits[n, t] = a[n, t] @ weight.T + bias, at every position t" } ], "output": "logits [N, seq_len, n_classes] at every position; the prediction at a position is the argmax over classes; only OUT positions carry an answer", "reference_forward": "forward.py: forward(params, x_int64 [N, seq_len]) -> float32 logits [N, seq_len, n_classes]" }