JackHsieh/luna-reason-only.k-8.statml-arxiv-llama32.llama32-ids.kv-tags-explained

luna-reason-only.k-8.statml-arxiv-llama32.llama32-ids.kv-tags-explained

Tokenized, tag-wrapped form of `JackHsieh/luna-reason-only.k-8.statml-arxiv-llama32`. Each thought is wrapped as

<|reserved_special_token_1|>
This is a hint about a span that appears later in this document. KEY is the text immediately before that span; VALUE is a note about what might come next.
KEY: <last 8 prefix tokens>
VALUE: <thought>
<|reserved_special_token_2|>

and stored both as text (thought_text) and as Llama 3.2 token ids (input_ids).

Longest thought: 547 tokens — a training run's max_thought_length must be at least this.

Intended to be PREPENDED to the document rather than interjected between prefix and suffix: the key names the span the value is about. The key is the 8 document tokens ending at the cut (input_ids[chunk_start_index - 8:chunk_start_index]), spliced as ids rather than re-tokenized.

Provenance

Source thoughts: JackHsieh/luna-reason-only.k-8.statml-arxiv-llama32 — prefix-only reasoning about the next 8 Llama-3.2 tokens of stat.ML arXiv LaTeX, generated by gpt-5.6-luna (reasoning.effort="none") from the last ≤1024 tokens before each cut. Documents: JackHsieh/statML-arxiv-40M-20M-llama32. Tokenizer: meta-llama/Llama-3.2-3B, add_special_tokens=False; tag ids and any document ids are inserted explicitly, so no BOS is introduced and no token is round-tripped through text.

Schema

All source columns, plus/with:

Token counts (input_ids length, tags included)

Nothing is trimmed: 163,893 thoughts exceed 384 tokens, so a run with a shorter max_thought_length must trim them (at a sentence boundary) or drop those chunks.