Drummer / sentence completion

The JSON is the received result, without a rewritten answer. Token IDs preserve generated tokens even when readable decoding is incomplete.

This is the 126-million-parameter foundation experiment, trained for 100 million next-token targets. That is a count of training exposures, not 100 million unique facts or sentences. The model has no conversation history on this page.

The trained weights and tokenizer are original to this project. Training text comes from a bounded generated-text subset of SmolLM-Corpus’s Cosmopedia v2, produced with Mixtral-8x7B-Instruct-v0.1. The generator was an offline source of training text, not a runtime assistant. Private user conversations, clinical documents and the project’s separate teacher-chat corpus are not inputs to this foundation run.

Contains information from SmolLM-Corpus, by Loubna Ben Allal, Anton Lozhkov, Guilherme Penedo, Thomas Wolf and Leandro von Werra (Hugging Face, 2024), under the ODC Attribution License. Database licensing does not certify rights to redistribute every underlying passage.

Systemic functional linguistics supplies a way to examine meaning: who does what and in which circumstances; requests, commitments and uncertainty; and how references connect a conversation. Preserving expressed feelings includes preserving whose feelings they are—not claiming that the model feels them.

This foundation checkpoint was trained without SFL labels. It is a starting control, not evidence that SFL improves model performance. SFL-guided conversation teaching, tool use and communicative compression are separate experiments. A plausible completion here does not establish those abilities.

A request allows up to 128 generated tokens and ten seconds of generation, within an approximately twenty-second whole-request limit. Only one generation is admitted at a time; a busy service asks you to try again yourself. There are no automatic retries.

Your prefix is sent to this service to generate a continuation. Do not submit sensitive text. Neither this page nor the inference service saves prompts or responses. They stay in memory for the request and current display, with no saved conversation history, local storage or training enrollment. Nothing you type retrains the model. Pageview analytics and infrastructure connection metadata are separate from model text.

Page analytics use a fixed page identity. Input text, output text and token IDs are not included in analytics events, page titles or URLs. The text and raw-result copy buttons copy only when you ask.

Public explanations and model identities are selective, reviewed disclosures—not publication of full private logs. Credentials, personal material and private or third-party records are excluded from those disclosures. Training-source attribution is retained separately.

An ending token establishes that generation ended, not that the answer is accurate. Partial text is labelled as partial; invalid bytes can appear escaped in diagnostic output. Original failures are shown, never repaired into an answer.

This unfinished model can invent facts, produce offensive text, repeat itself and lose the topic. It is an experiment, not advice or a dependable assistant.

I am sharing this stage while it is still imperfect. The foundation checkpoint has learned some sentence structure, but direct tests still produce repetition and topic drift. Earlier controlled-language rewriting and tiny dialogue checkpoints did not provide usable open-ended chat; those experiments remain separate from this demo.

For example, the prefix “The gardener noticed that” produced this opening verbatim: “ the soil was too strong and the soil was too strong.” It continued repeating and reached the 128-token limit without an ending token. This is an exploratory observation, not a benchmark result.

The demonstration exposes the frozen 100-million-target checkpoint, not the separate larger training schedule. Checkpoint changes will be explicit. Good-looking completions do not establish an advantage over other models.

Future process excerpts will preserve the original wording, mistakes and revised decisions. Any removed credentials, personal or third-party private material, or unapproved unpublished work will be marked as redacted. Excerpts will not be presented as the entire unedited private record.