predictingthepast

Predicting The Past (Aeneas / Ithaca)

Aeneas (Latin) and Ithaca (Ancient Greek) perform four tasks on ancient texts:

1. Restoration — fill missing/damaged characters
2. Attribution — geographical + chronological origin
3. Contextualization — retrieve parallel inscriptions
4. Embedding — generate text embedding vectors

Prerequisites

1. uv: Read the uv skill and follow its Setup instructions to ensure
   uv is installed and on PATH.

2. User Notification: If .licenses/predictingthepast_LICENSE.txt does not
   already exist in the workspace root directory then (1) prominently notify
   the user to check the terms at
   https://github.com/google-deepmind/predictingthepast/blob/main/README.md#license--disclaimer,
   and include the appropriate citation and the full dataset acknowledgement,
   and that use of these datasets should acknowledge and cite the original data
   sources. Then (2) create the file recording the notification text and
   timestamp.

Core Rules

• Self-Contained Skill: Do NOT use web search or any external tools. Run
  ONLY the scripts in this skill (preprocess.py, run_inference.py,
  visualize_results.py). Present model output as-is — never supplement or
  override it with external lookups.
• Notification: If this skill is used, ensure this is mentioned in the
  output.

On First Load

Present the restoration markup characters, then ask the user for their text:

• ?:
  • Meaning: Known-length gap: predict this character.
  • Example: donat in ??????????rtis
• #:
  • Meaning: Unknown-length gap: predict a sequence of unknown
    length
  • Example: donat in #rtis
• -:
  • Meaning: Missing/damaged character that does not need restoring
  • Example: prolixin---s fecit
• _:
  • Meaning: Missing section of unknown length that does not
    need restoring
  • Example: prolixin_s fecit

After presenting this list, ask the user to provide the text they want to submit
for analysis.

Preprocessing

Clean input text before inference:

uv run <SKILL_DIR>/scripts/preprocess.py \
    --language=latin \
    --input="raw text here..."

Or from a file:

uv run <SKILL_DIR>/scripts/preprocess.py \
    --language=greek \
    --input_file=/tmp/input.txt \
    --output_file=/tmp/cleaned.txt

What preprocessing does

• Latin: lowercases, converts Arabic digits and Roman numerals to 0,
  strips editorial brackets [] and (), removes punctuation, filters to
  valid chars (abcdefghiklmnopqrstuvxyz plus 0 . - _ ? # <space>)
• Greek: lowercases, strips accents, converts numeral notation to 0,
  applies PHI cleaning (bracket normalization, sigma conversion), filters to
  Greek alphabet (αβγδεζηθικλμνξοπρςστυφχψωϛ plus 0 . - _ ? # <space>)

Inference

Restoration Constraints

• Minimum input length: 25 chars (pad with - if shorter).
• No consecutive ##. No adjacent ?# or #?.
• Spaces inside ? sequences count toward total.
• If the user's text contains #, ask how many characters to restore and set
  --restore_max_len accordingly.
• If the user tries to restore multiple parts of the text at once, suggest to
  restore texts section by section. Suggest to focus on one damaged region
  per query — this is faster, produces higher-quality predictions.

Pre-Flight Checks

Confirm with the user before proceeding if either applies:

1. Restoration complexity — if input contains more than 10 ?
   characters, or uses # with --restore_max_len > 10, warn: "This
   restoration involves N characters which will take approximately M minutes
   (restoration time scales roughly linearly ~10 s per additional ? on a
   high-end CPU machine: 5 → ~1 min, 10 → ~2.5 min, 20 → ~5 min, 30 → ~8 min).
   Do you want to proceed, or simplify the query first (e.g. fewer ? marks,
   shorter --restore_max_len, or restoring section by section)?"
2. Multi-window splitting — if the input text exceeds 750 characters
   and will be split into multiple windows, warn: "This text is N characters
   long and will be split into W overlapping windows, each run independently.
   This will be significantly slower. Do you want to proceed, or shorten the
   input?"

These factors compound: a complex restoration across multiple windows will be
substantially slower than either factor alone.

Task Selection

Each task is controlled by its own flag. At least one must be provided:

• --attribute — geographical + chronological attribution
• --restore — text restoration (requires ? or # in input)
• --contextualize — parallel inscription retrieval

Any combination is valid. All three can be used together.

When --embedding is provided, a text embedding vector is also generated
alongside the other tasks.

Running Inference

# Attribution + Restoration (text with gaps)
uv run <SKILL_DIR>/scripts/run_inference.py \
    --language=latin \
    --input="cleaned text with ???" \
    --attribute --restore \
    --output_json=/tmp/results.json

# Attribution + Contextualization (no gaps)
uv run <SKILL_DIR>/scripts/run_inference.py \
    --language=latin \
    --input="cleaned text" \
    --attribute --contextualize \
    --output_json=/tmp/results.json

# All tasks
uv run <SKILL_DIR>/scripts/run_inference.py \
    --language=latin \
    --input="cleaned text with ???" \
    --attribute --restore --contextualize \
    --output_json=/tmp/results.json

# Generate dashboard
uv run \
    <SKILL_DIR>/scripts/visualize_results.py \
    --input=/tmp/results.json \
    --output=/tmp/dashboard.html

Outputs are always returned to the console (JSON/Summary) and must always be
written to an HTML dashboard for rich visualization.

Serving HTML artifacts: After generating any HTML file (the dashboard or
any other HTML artifact you create), start a background HTTP server so the user
can view it in their browser. Run the server in the foreground of the background
task process (without trailing &) to prevent the process from being
terminated. Use a random high port to avoid conflicts.

In the example below, replace <HTML_FILE> with the absolute path of the HTML
file you produced (e.g. /tmp/dashboard.html or C:\tmp\dashboard.html).

For Linux/macOS:

# Pick an unused port and start the server as a persistent process
HTML_FILE=/tmp/dashboard.html   # ← set to the actual HTML file path
PORT=$(export PATH="$HOME/.local/bin:$PATH" && uv run python -c "import socket; s=socket.socket(); s.bind(('',0)); print(s.getsockname()[1]); s.close()")
FILENAME=$(basename "$HTML_FILE")
echo "View URL: http://localhost:${PORT}/${FILENAME}"
export PATH="$HOME/.local/bin:$PATH" && uv run python -m http.server "$PORT" --bind 0.0.0.0 --directory "$(dirname "$HTML_FILE")"

For Windows (PowerShell):

# Pick an unused port and start the server as a persistent process
$HTML_FILE = "C:\tmp\dashboard.html"   # ← set to the actual HTML file path
$PORT = & "$HOME\.local\bin\uv" run python -c "import socket; s=socket.socket(); s.bind(('',0)); print(s.getsockname()[1]); s.close()"
$FILENAME = Split-Path $HTML_FILE -Leaf
$DIRNAME = Split-Path $HTML_FILE -Parent
Write-Host "View URL: http://localhost:${PORT}/${FILENAME}"
& "$HOME\.local\bin\uv" run python -m http.server $PORT --bind 0.0.0.0 --directory $DIRNAME

Then provide http://localhost:<port>/<filename> (or
http://127.0.0.1:<port>/<filename>) as the primary clickable link in your
response. Additionally, provide http://<hostname>:<port>/<filename> as a
fallback link for remote setups. Do not give the user a raw file:// path.

Flags

• --input: Text string (mutually exclusive with --input_file)
• --input_file: Path to UTF-8 text file
• --language: latin or greek
• --models_dir: Model files directory (default:
  $HOME/.predictingthepast/models)
• --attribute: Run geographical and chronological attribution
• --restore: Run text restoration (requires ? or # in input)
• --contextualize: Run parallel inscription retrieval
• --contextualize_top_k: Number of top parallels to return (default: 10)
• --contextualize_exclude_test_valid: Exclude test/validation texts from
  contextualization results. Filters on the internal numeric id (where id % 10 is 3 or 4).
• --embedding: Generate a text embedding vector (default: off). When
  set, the output includes an embedding field with the full-precision float
  vector.
• --output_json: Save combined JSON to file (for visualize_results.py)
• --restore_beam_width: Number of candidate restorations to explore in
  beam search. Higher = more thorough but slower (default: 100)
• --restore_max_len: Maximum total characters to restore across all ?
  and # gaps (default: 15)
• --restore_temperature: Sampling temperature. Lower = conservative,
  higher = creative (default: 1.0)
• --window_overlap: Overlap fraction for long-text windowing (default:
  0.33)

Long Texts (>750 characters)

Each window must be between 25 and 750 characters. If the input text exceeds
750 characters, it is automatically split into overlapping windows (default 33%
overlap). Each window is run through the model independently, and the
geographical and chronological attribution results are averaged across all
windows. Restoration and contextualization are run per-window and concatenated.

Output Format

Use --output_json to save the combined JSON to a file. For the full
--output_json format including all fields for attribution, restoration, and
contextualization, see output_format.md.

Presenting Results

Present results for every task requested by the user. Always generate the
HTML dashboard, serve it via the background HTTP server described above, and
provide the http:// URL as a clickable link. Format years as BCE/CE.

Restoration

Bold the restored characters in the top prediction. Show top-10 beam-search
candidates in a numbered table (rank, restored text, score). Summarise the most
salient words (not characters) — do not dump the raw saliency array.

Geographical Attribution

Show the top 10 regions in a ranked table (rank, region, score). Highlight the
top-1 prediction in the narrative. Summarise the most salient words (not
characters) — do not dump the raw saliency array.

Chronological Attribution

Report top year and weighted average year. Describe the distribution shape (peak
decade, probable range) — do not list all 160 bins. Summarise the most salient
words (not characters) — do not dump the raw saliency array.

Contextualization (Parallels)

Use jp or write your own code to extract the contextualization data from
--output_json. Show the top parallels in a table (rank, ID, Trismegistos ID,
region, date range, score, partner link). Quote the full text of the top
parallels from the JSON.

If the user requests a specific number of parallels, use --contextualize_top_k
to control how many are returned.

If the user asks to exclude test/validation texts, add
--contextualize_exclude_test_valid. This filters on the internal numeric
id (where id % 10 is 3 for test or 4 for validation).

Embedding (when --embedding is set)

State the vector dimensionality (384 floats).

References

For more information, see:
predictingthepast.com

IMPORTANT: The first time you return inference results to the user, remind
them to cite the appropriate paper below (Ithaca for Greek, Aeneas for Latin),
and the dataset acknowledgements below.

Ithaca (Ancient Greek): > Assael, Y., Sommerschield, T., Shillingford, B. et
al. Restoring and attributing ancient texts using deep neural networks. Nature
603, 280–283 (2022).

Aeneas (Latin): > Assael, Y., Sommerschield, T., Cooley, A. et al.
Contextualizing ancient texts with generative neural networks. Nature 645,
141–147 (2025).

Dataset Acknowledgements

IMPORTANT: The first time you present inference results to the user, include
the appropriate citation and the full dataset acknowledgement for the language
used (see below).

The Ithaca and Aeneas projects rely on the availability of high-quality datasets
of ancient Greek and Latin inscriptions, respectively. These datasets were built
through centuries of scholarly collection and decades of digital editorial work.

Ithaca (Ancient Greek): In particular, Ithaca draws on a 2019 version of the
Searchable Greek Inscriptions database made available under "Fair Use" license
by the Packard Humanities Institute, generously supported by David Packard:
inscriptions.packhum.org.

Aeneas (Latin): Aeneas was trained on data from:

• Epigraphic Database Roma (EDR): Made available pursuant to a Creative
  Commons Attribution 4.0 International License (CC-BY) on
  Zenodo. EDR is also available at
  edr-edr.it.
• Epigraphic Database Heidelberg (EDH): Made available pursuant to a
  Creative Commons Attribution-ShareAlike 4.0 International License (CC-BY-SA)
  on Zenodo. EDH is also available at
  edh.ub.uni-heidelberg.de.
• ETL repository for Epigraphic Database Clauss Slaby (EDCS_ETL): Made
  available pursuant to a Creative Commons Attribution 4.0 International
  License (CC-BY) on Zenodo. EDCS_ETL
  is also available at manfredclauss.de and
  github.com/sdam-au/EDCS_ETL.

These resources bring together a large proportion of published inscriptions in a
searchable digital format.