{
  "architecture": "PLE",
  "format": {
    "magic_ascii": "PLE\\0",
    "magic_le_hex": "0x00454c50",
    "version": 1,
    "header_bytes": 56,
    "flags": {
      "TIED_HEAD": false
    }
  },
  "config": {
    "input_vocab": 8057,
    "output_vocab": 854,
    "d_model": 128,
    "n_layers": 6,
    "n_heads": 4,
    "ffn_hidden": 384,
    "ple_dim": 128,
    "seq_len": 128,
    "rope_theta": 10000.0
  },
  "vocabulary": {
    "note": "asymmetric: the model reads a BPE vocabulary and writes a separate, smaller output alphabet of 854 classes, which includes punctuation and specials, so the output head is its own tensor rather than a view of the embedding",
    "bpe_entries": 7338,
    "word_rows_appended": 719,
    "input_vocab_total": 8057,
    "output_classes": 854,
    "output_classes_reusing_a_bpe_id": 135,
    "input_encoding": "ASCII only; non-ASCII input is refused rather than encoded differently from the reference tokenizer"
  },
  "parameters": {
    "ple_table": 6187776,
    "tok_emb": 1031296,
    "output_head": 109312,
    "core": 1575424,
    "total": 8903808
  },
  "quantization": {
    "weights": "int4",
    "group_size": 128
  },
  "runtime": {
    "weights_staged_int8": "per-position core and the untied output head, in PSRAM",
    "activations": "int8, quantized per staged matvec",
    "sram": "float scratch buffers and RMSNorm vectors",
    "psram": "staged core, staged head, KV cache",
    "flash": "PLE table and token embedding, memory-mapped"
  },
  "measured": {
    "unit": "an output piece is one emitted class; punctuation is a class. Over the benchmark prompt set, 253 pieces render as 213 readable words.",
    "serial_only": {
      "ms_per_piece": 60.25,
      "pieces_per_second": 16.6,
      "readable_words_per_second": 13.97
    },
    "with_oled": {
      "ms_per_piece": 88.83,
      "pieces_per_second": 11.26,
      "readable_words_per_second": 9.48,
      "added_ms_per_piece": 28.58
    },
    "ms_per_forward": 49.6,
    "method": "scripts/benchmark_device.py, eight fixed prompts, two passes per mode, spread under 0.1%, same weights and build switches apart from the display, generated text identical in both modes",
    "common": {
      "staged_tensors": 44,
      "scratch_sram_bytes": 20940,
      "norm_vectors_in_sram": 20,
      "sram_fallbacks": 0,
      "build": "-O3",
      "date": "2026-08-02"
    }
  },
  "device_fingerprint_fnv1a": "e602146b",
  "files": {
    "model.bin": {
      "sha256": "1359a1cb74de4143d630c2c192990de814cd47255bcdfa9cc135f07ef0a39fc4",
      "bytes": 4600186,
      "license": "MIT"
    },
    "tokenizer.json": {
      "sha256": "0ad085811c949f35c5f5f15b555f2ff2d46ec1ec94a1650552416d06aaa19ee2",
      "bytes": 491735,
      "license": "MIT"
    },
    "vocab.json": {
      "sha256": "5a16d6224abf03265d69ebcccf121c8f8d2c222bfedd8274acbc0bbbe13e4eb7",
      "bytes": 50494,
      "license": "MIT"
    },
    "layout.json": {
      "sha256": "15036c5ee2b23b9b35404ef6422cb788bbede8cf2c269bae35d4dbf9a48a0b90",
      "bytes": 5142,
      "license": "MIT"
    }
  },
  "weights_license": "MIT",
  "tokenizer_license": "MIT",
  "source_checkpoint": {
    "name": "v5-L6-s0.pt",
    "distributed": false,
    "sha256": "bad2cfb6627dbde6893b11cd7e773ec96f1d8c36bf6d3a5d36845f9edc5f0ffb"
  },
  "training_data": {
    "dataset": "private question and answer corpus written for this project",
    "redistributed": false,
    "note": "the corpus, the checkpoints and the evaluation sets are not distributed. The weights and the inference assets in this repository are, which is why the vocabulary files are here: without them the output classes cannot be turned back into words or fed forward."
  },
  "hardware": {
    "board": "ESP32-S3",
    "flash_mb": 16,
    "psram_mb": 8,
    "sram_kb": 512,
    "model_partition_offset": "0x110000"
  }
}