Pretrained tools#

Install tensorcode[tools] as shown in the quickstart. Hugging Face is the preferred host. Tools construct their own encoders, workspace and prediction/decoding components; callers do not supply a separate model.

The release records contain pinned revisions and remote-loading checks. Model cards include evaluation records, training data provenance and limitations. Validation reports the baselines and ablations: these checkpoints do not establish a consistent benefit from the recurrent slot workspace or general cognitive competence.

The listed Chatbot checkpoints and the cognitive Investigator use the earlier unbounded workspace update. Reproduce them with source commit 6607a8b; they are not compatible with the current bounded-update architecture. Matching replacement weights have not qualified yet.

Hotpot Investigator and Decision have configuration-only refreshes: Investigator revision 1bc225917c3646fcb9702df91ff5e445846c1dc7 and Decision revision 59b9f8d2c09e4d3179f601d54efafed1da45648d. These explicitly record three existing defaults. Published loading, weight bytes, full model state and three supplied- candidate probe receipts match the historical runtime exactly. Their original performance scope is unchanged; this is not new training or cognitive qualification. Original pins in the historical release record require commit 6607a8b. Planner and Scene manifests show no constructor-default drift in a configuration-only audit; that audit does not establish current full-loading or prediction parity.

Current Chatbot configuration records memory_update="relative_rms_bounded". Each workspace update is normalized against its example's unmasked encoder RMS and uses a bounded gate. This controls update magnitude, with floating-point rounding tolerance; it does not establish useful learned reasoning. Loading rejects configuration drift, including nested defaults and JSON value types, before applying weights.

To load the refreshed ranking checkpoint on current source:

from tensorcode.tools.investigator import Investigator

model = Investigator.from_pretrained(
    "jacob-valdez/tensorcode-investigator-hotpot-001",
    revision="1bc225917c3646fcb9702df91ff5e445846c1dc7",
)

Construct, call, save and reload#

from tensorcode.tools.investigator import Investigator

# Fresh random parameters: this demonstrates the current artifact lifecycle.
model = Investigator({
    "vocabulary": ["which", "document", "describes", "sky", "blue", "kettle", "water"],
    "dimensions": 32,
})
result = model({
    "question": "Which document describes the sky?",
    "evidence": [],
    "hypotheses": [
        {"id": "document-a", "text": "The sky appears blue during the day."},
        {"id": "document-b", "text": "A kettle boils water."},
    ],
})
print(result["selected_id"])
model.save_pretrained("./my-investigator")
restored = Investigator.from_pretrained("./my-investigator", local_files_only=True)

This fresh model has not been trained; its selected candidate demonstrates the interface and is not evidence of ranking quality. The complete model and its encoding assets are saved locally. Session history and optimizer state are saved separately; see training and tool sessions.

from_pretrained accepts revision, cache_dir, local_files_only, token, and device. A local directory loads without contacting the Hub. Tool(config) initializes fresh parameters without downloading anything. Explicit from_foundation methods bootstrap inherited perception/language weights and a new workspace for training; they are not equivalent to loading a trained TensorCode checkpoint.

Reproduce the earlier experimental cognitive Chatbot#

Use TensorCode commit 6607a8b for this checkpoint. It includes the loader correction that initializes fresh memory after restoring weights and retains the architecture used for the recorded evaluation. This complete checkpoint owns the generator, verifier, retrieval encoder and realizer. It is suitable for inspecting and training the pipeline; its final 32-question evaluation produced only one correct answer, one non-answer and 30 abstentions. The validation report explains that limitation.

from tensorcode.tools.chatbot import Chatbot

bot = Chatbot.from_pretrained(
    "jacob-valdez/tensorcode-chatbot-cognitive-experimental-001",
    revision="8836ba59275dc6d8ceeb04462b4191beb9813452",
)
response = bot({
    "question": "How did the service recover?",
    "evidence": [{"id": "incident", "source_id": "report:17",
                  "text": "The service recovered after reconnecting the database."}],
})
print(response)  # May abstain; inspect the model judgments in bot.last_result.
bot.new_episode()  # Retains source memory, clears the current interpretation.
bot.save_session("./session.json")

The complete Investigator component is separately available at jacob-valdez/tensorcode-investigator-cognitive-experimental-001, revision a5ee35f6c644fa13d4bda5850ea37eb32ec3f6d2. Its investigate method generates and screens hypotheses without the outer Chatbot realizer. The Scene language checkpoint is pinned at 6aa55691bbdab5f49edede24ae0a71a673e2cc26; see the visual interpretation interface.

Publish a fine-tuned model#

model.push_to_hub(
    "your-account/my-investigator",
    model_card="# My Investigator\n\nDescribe data, training, evaluation and limits here.\n",
)

Authenticate through the Hugging Face CLI or token configuration before publishing. Publication includes model assets and the card, not conversation/session or optimizer state. Supply actual evaluation and provenance in the card; the default card identifies the architecture and makes no performance claim.