Aparecium

Aparecium reconstructs short English text from the full token-state matrix produced by sentence-transformers/all-mpnet-base-v2. A nonlinear projection maps each token state to the encoder's original vocabulary. Reconstruction receives the matrix without requiring the source text.

The result is an approximate reconstruction. Names, numbers, dates, conditions, and actions can be omitted or changed. Use the output to inspect represented content, and check important claims against source material. It is not verified news or a basis for unattended trading decisions.

The input is AutoModel.last_hidden_state, including its first and last framing rows, from the pinned encoder and tokenizer described below. Pooled sentence vectors, including 768-element vectors from SentenceTransformer.encode, cannot be used. A stored pooled vector does not contain the full token states needed by this model.

Install

pip install --pre aparecium

The default model is SentiChain/aparecium-matrix. Its weights and tokenizer are downloaded on first use and cached locally.

Reconstruct a stored matrix

import numpy as np
from aparecium import Aparecium

matrix = np.load("matrix.npy", allow_pickle=False)
model = Aparecium()
result = model.invert_matrix(matrix)

print(result.text)
print(result.token_ids)
print(result.input_rows)

invert_matrix uses only the supplied states and optional attention mask. It does not re-embed text or retrieve source material.

The input contract is:

• Token states with shape (sequence_length, 768), containing 3 through 256 rows. Preserve native float32 encoder output when storing states.
• States from the pinned encoder and tokenizer, without pooling, row normalization, quantization, or dimensionality reduction.
• The first and last visible rows are the encoder's beginning and end framing positions. The decoder removes these positions before returning token IDs.
• An optional binary attention mask of the same sequence length. Visible rows must form a prefix followed by any right padding, with at least three visible rows.
• Finite real values only. Numeric inputs are converted to float32; overflow is rejected. Other embedding models are incompatible even if their output width is 768.

To reconstruct up to 32 matrices, pass a batch array or a list of matrices. Lists can contain different sequence lengths:

results = model.invert_matrices([matrix, matrix])
for result in results:
    print(result.text)

Use attention_masks= when a batch contains padded rows. Matrix width and shape checks cannot establish encoder provenance: record the encoder identifier, revision, and preprocessing alongside stored states.

Produce compatible matrices

Use the encoder and tokenizer at revision e8c3b32edf5434bc2275fc9bab85f82640a19130. Save the complete final hidden states rather than a pooled or normalized sentence vector:

import numpy as np
import torch
from transformers import AutoModel, AutoTokenizer

encoder_id = "sentence-transformers/all-mpnet-base-v2"
encoder_revision = "e8c3b32edf5434bc2275fc9bab85f82640a19130"
tokenizer = AutoTokenizer.from_pretrained(encoder_id, revision=encoder_revision)
encoder = AutoModel.from_pretrained(
    encoder_id, revision=encoder_revision, dtype=torch.float32
).eval()

inputs = tokenizer(
    "Heavy rain forced the city to close several roads.",
    return_tensors="pt",
    truncation=False,
)
if not 3 <= inputs["input_ids"].shape[1] <= 256:
    raise ValueError("Text must produce 3 through 256 token rows, including framing.")

with torch.inference_mode():
    matrix = encoder(**inputs).last_hidden_state[0].cpu().numpy()

np.save("matrix.npy", matrix, allow_pickle=False)

For a batch producer, retain each attention mask and keep only its visible prefix or pass the mask with its padded matrix. Do not truncate overlong source text silently: reconstruction can recover only the content represented by the stored states.

Text convenience methods

invert_text creates compatible states from one string before reconstructing them. invert_texts accepts up to 32 strings. These are round-trip convenience methods; stored-matrix reconstruction uses invert_matrix or invert_matrices.

result = model.invert_text("Heavy rain forced the city to close several roads.")
print(result.text)

These methods use the pinned AutoTokenizer and AutoModel.last_hidden_state, with no pooling. Text that produces more than 256 rows is rejected. The encoder weights are downloaded separately when a text convenience method first needs them.

Loading and devices

model = Aparecium(
    model_id="SentiChain/aparecium-matrix",
    revision=None,
    model_path=None,
    device=None,
    local_files_only=False,
    token_batch_size=256,
)

CPU is the default device. Set device="mps" explicitly to use a supported Apple GPU. The runtime supports CPU and MPS, and an unavailable requested device raises an error. token_batch_size accepts 1 through 256 and controls how many state rows are projected together; it does not alter the sequence-length or matrix-count limits.

Set revision to a Hub commit for reproducible model loading, or use model_path for a local artifact directory. local_files_only=True prevents downloads. Matrix reconstruction then needs the complete local model artifact; text convenience methods also need the pinned encoder files in the local cache.

Command line

aparecium --matrix-file matrix.npy
aparecium --matrices-file matrices.json --json
aparecium "Heavy rain forced the city to close several roads."
echo "Heavy rain forced the city to close several roads." | aparecium --json

The equivalent module command is python -m aparecium. Choose a matrix file, a matrices file, or text. Without a file or positional text, the command reads text from standard input.

--matrix-file accepts a two-dimensional .npy file, a JSON matrix array, or a JSON object containing matrix and an optional attention_mask. --matrices-file accepts a three-dimensional .npy file, a JSON list of matrices, or a JSON object containing matrices and optional attention_masks. JSON supports matrices with different sequence lengths. Input files are limited to 16 MiB.

--json returns the result fields above. Batch JSON has a results array. Shared options are --model-id, --revision, --model-path, --device, --local-files-only, and --token-batch-size.

HTTP service

pip install --pre 'aparecium[server]'
aparecium-serve --host 127.0.0.1 --port 8000

The equivalent module command is python -m aparecium.service. The service accepts the shared model options above. APARECIUM_LOCAL_FILES_ONLY=true also enables offline loading.

GET /health reports readiness and the input contract. POST /invert accepts either matrix with optional attention_mask, or matrices with optional attention_masks. Supply one of those two forms and no other fields. The service does not need source text.

This example sends a complete saved matrix:

import json
from urllib.request import Request, urlopen
import numpy as np

matrix = np.load("matrix.npy", allow_pickle=False)
request = Request(
    "http://127.0.0.1:8000/invert",
    data=json.dumps({"matrix": matrix.tolist()}).encode(),
    headers={"Content-Type": "application/json"},
    method="POST",
)
with urlopen(request) as response:
    print(json.load(response))

A single request returns text, token_ids, and input_rows; a batch request returns a results array. Unknown fields, pooled vectors, invalid masks, and nonfinite values are rejected. Request bodies are limited to 16 MiB. Invalid inputs return HTTP 422, oversized bodies return 413, and an unavailable model returns 503.

Quality and limitations

On a held-out collection of 20 historical crypto event families with three posts per event, all 20 reconstructed event accounts retained a useful central claim. Under two source-fidelity assessments, 18 were useful without a critical assertion and two contained critical assertions. These are judgment-based assessments, not independently human-validated accuracy estimates or verified market outcomes. The collection was balanced across bullish, bearish, mixed, and unclear source context rather than sampled from a live news stream.

Token recovery and broad event usefulness do not guarantee safe financial detail. In recorded cases, about 2,000 BTC became about 2,000,000 BTC, and a forecast horizon of 2030 became 2036. Forecast bounds, secondary quantities, and conditions can also disappear. Combining multiple reconstructed posts can repeat a shared error.

The model is intended for short English content from the specified encoder. Exact quotation recovery, long documents, other languages, trading profitability, and unattended source replacement have not been established. See the benchmark details and model card for training provenance, field-level results, and measurement limits.

Input compatibility

Integrations that store pooled sentence vectors must capture full token states to use this model. Replace pooled-vector calls with invert_matrix or invert_matrices, preserve framing and attention masks, and update service payloads to matrix or matrices. Pooled checkpoints and vector APIs are unsupported.

License and attribution

Aparecium code and the learned projection are licensed under the MIT License. The included original MPNet vocabulary weights and tokenizer retain their Apache-2.0 terms and notices. Training sources retain their own terms; source text is not redistributed with the model. See ATTRIBUTIONS.md.