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.

Current revisions and historical pins#

All eight checkpoints load at main with TensorCode 0.4.0a4. The two Chatbots and the cognitive Investigator were retrained on the current bounded workspace update and published as new commits on their existing repositories. Their earlier unbounded-update revisions are unchanged and still load with source commit 6607a8b only.

To reproduce a historical result, install TensorCode from source commit 6607a8b and pass the historical pin as revision. On that runtime the three replaced pins still load and reproduce the recorded answers checked (three held-out generations, the Franco answer and its Investigator receipt); the cognitive receipt additionally records verification_scope and a different model_provenance hash than the original 50f170e run.

The replacements are measured against the recorded results of the revisions they replace (Chatbot, cognitive Chatbot and Investigator):

• Hotpot Chatbot: the original recipe, data and seed on current source. Held-out exact match 31/64 and token F1 0.6171 (previously 30/64 and 0.6015). Bypassing the workspace gives 30/64 and a lower cross-entropy, so no workspace benefit is claimed.
• Cognitive Chatbot: retrained generator and realizer; verifier, ranker and retrieval encoder reused byte-for-byte. On the same 32 known questions it answers 3 and abstains on 29: two answers are correct and one is a wrong fact that the verifier accepted (previously one correct answer and one circular non-answer). The generator's foundation learning rate (5e-5 instead of 3e-5) was chosen after the test scores of all variants had been seen; the model card discloses this.
• Cognitive Investigator: the component extracted from that Chatbot, with a bitwise-equal state dict and equal investigate receipts on all 32 cases. It still has no independent answer benchmark.

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 both Scene checkpoints load at their original revisions on current source; a fresh-cache load and prediction check found Python and TypeScript receipts equal (strings exactly, numbers within 1e-4 absolute plus 1e-3 relative) on sample inputs. That check does not re-measure their task performance.

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.

Load the experimental cognitive Chatbot#

This complete checkpoint owns the generator, verifier, retrieval encoder and realizer. It is suitable for inspecting and training the pipeline. On 32 reused HotpotQA questions with oracle passages it answered three (two correct, one wrong fact) and abstained on 29. The validation report explains these limits.

from tensorcode.tools.chatbot import Chatbot

bot = Chatbot.from_pretrained(
    "jacob-valdez/tensorcode-chatbot-cognitive-experimental-001",
    revision="6fc386fbfe992b8fb02c9049d4a761ed51c54132",
)
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 267b4f00d63905f0d8b9b87ba3542896d7c5e9e7. 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.