zachnorton03/vintage-sft-robustness

Vintage SFT Robustness

Supplementary SFT rows for a pre-1930 conversational model. 7,338 rows across five routes. Kept deliberately separate from the graded synthetic dataset, whose rows are answers lifted verbatim from period prose and scored by a judge; these are constructed, ungraded, and would muddy that provenance guarantee if mixed in.

Why it exists

A model pretrained on pre-1930 books and finetuned on question-and-answer pairs behaves correctly right up until the input stops looking like a question. Then it falls off the finetuned distribution, falls back on base-corpus behaviour — book prose, which carries almost no end-of-sequence token — and continues until the token budget runs out.

Observed failures that motivated each route:

input	what happened	route
Hello	paragraphs of narration	conversation_qa
xakhavjba	same	unparseable_qa
a question with the ? dropped	same	train-time noise (below)
What is a nedle?	answered something else	typo_qa
What year is it?	wrong or evasive	era_qa
any multi-turn exchange	no model of taking turns	conversation_multiturn

The last one has a structural cause worth recording. The graded dataset was built by scoring excerpts on region_quality, which rewards mean sentence length. Dialogue turns are 2–6 words, so quoted speech scores near zero on a quarter of that metric and lands below the 0.70 keep floor. Measured on the source corpus:

excerpts with no dialogue   0.741 mean prose score   (passes)
excerpts with dialogue      0.62–0.64                (fails)

The SFT data is therefore ~98% narration. The model trained on books, but on the parts of books that are not people talking.

Contents

route	rows	median tokens/row	what it covers
conversation_qa	3,500	15	greetings, farewells, thanks, acknowledgements, odd openers (Praise God!), vague references (Does he so?)
unparseable_qa	2,000	18	gibberish and cut-off sentences
typo_qa	1,200	30	one word mangled hard → You mean needle, I take it?
era_qa	220	17	year (147), anachronisms (31), identity (25), place (17)
conversation_multiturn	418	86	genuine two-party dialogue, 2,044 turns
total	7,338		~142k tokens per pass

unparseable_qa at 2,000 rows: fragment 655, multi_token 365, single_nonword 349, consonant_run 193, long_run 99, alnum_noise 93, micro 73, mixed_case 69, keyboard_mash 64, repeated_char 32, bare_punctuation 8. The last two saturate — their distinct value spaces are finite and rows deduplicate on the question.

Row count and token weight diverge sharply. These rows are short. In a curriculum where they are 3.4% of rows they can be 0.6% of tokens, so the gradient signal is far smaller than the row count suggests. Size the dose by tokens, not rows.

Single-turn rows carry question / answer; conversation_multiturn carries a conversations list of alternating {role, content}. All carry score: 100 — they are constructed, not graded, and must clear any curriculum grade threshold.

How each part was sourced

Three provenance classes, and only one carries anachronism risk.

Corpus-verbatim — lifted from a 988 MB pre-1930 excerpt corpus (452,591 excerpts):

• 90 deflection/social clauses, kept only when context-free (a terminator follows rather than a complement, so I don't know what you mean qualifies and ...by load factor does not)
• 5,000 whole sentences, truncated at a random word boundary to make fragments
• 4,000 short standalone utterances, used as odd openers
• all conversation_multiturn turns, from croqaz/vintage-conversations (Gutenberg novels, verified letter-by-letter against the source)

Procedural — random strings and mechanical mangling. Carries no register at all, so no anachronism risk: the gibberish classes, and the five typo operations (drop, swap, double, QWERTY-neighbour, phonetic).

Authored — 115 distinct strings written by hand in period voice. This is the entire anachronism surface. Everything else is corpus text or random characters.

 32  conversation openers ("What brings you", "What is the news with you")
 12  pronoun-specific asks ("Who is 'he'? We have not spoken of anyone.")
 12  year answers ("Nineteen hundred and thirty.", "1930, and well into it.")
 10  typo confirmations ("You mean {word}, I take it?")
  9  fallback acknowledgements, 9 identity answers, 8 closers,
  7  curious tails, 6 generic asks, 5 nature answers, 5 place answers
115  TOTAL

At 7,338 rows those 115 strings recur heavily — a bad one repeats thousands of times. They are also 115 lines, readable in five minutes, and worth reading before training on them.

Anachronism risk warnings

1. Corpus-verbatim protects against modern leakage, not against archaism. The corpus spans centuries. good morrow was the single most frequent greeting match and adieu the second most frequent farewell — both Shakespearean-to-Georgian, wrong for 1930. Filtering by the excerpt's year does not fix this: good morrow is more frequent in 1900s-published books than in 1800s ones, because publication year does not track register (the corpus holds reprints and historical fiction that are archaic on purpose). Curation by hand was the only reliable filter, and the blocklist is necessarily incomplete.

2. Frequency is not evidence of fitness. come again was the most common "invitation" match by a wide margin, precisely because the period sense is return, not repeat that — every sampled occurrence was astronomical recurrence or a parting pleasantry. It is blocklisted. Others of the same kind may survive.

3. conversation_multiturn skews Victorian, earlier than 1930. Sourced from Dickens, Eliot, Tolstoy in translation. This is defensible — the pretraining corpus is itself modal around 1850–1910 — but it is not 1930.

4. Dialect survives in user turns. Phonetic dialect (Know'd it yes'day aft'noon at tea-time) is rejected from assistant turns only. It remains in some user turns, on the reasoning that a visitor may type anything but the model should not answer that way.

5. Speaker attribution in the upstream dialogue set is not guaranteed. Its own card says so. Mechanical filters catch structural breaks, not semantic ones, so some exchanges will not cohere even though every turn is verbatim.

6. The user side is deliberately modern, and that is not a bug. hi, hey, what's up, ok appear as inputs on purpose: the visitor is a person at a keyboard today. Only the answers need to hold period register.

7. Constructed answers are not corpus-attested. The 115 authored strings are assembled in period voice but no period book contains them. They pass no anachronism filter because none was run on them.

Companion: train-time input noise

Not part of this dataset, but designed with it. The consuming trainer batters a share of user turns across the whole curriculum — 16 operations in five families: punctuation (including doubled Hello!!), case (HELlo, shout, leading-lower), spacing (Whatis, What is), character (transpose, dropped/doubled letter, QWERTY-neighbour), and word (dropped, repeated, apostrophe stripped). Seeded per epoch, so a row reads clean on one pass and battered on the next.

Answers are never touched. The reason is worth stating: a corrupted question is context the model reads; a corrupted answer is a target it learns to reproduce.

typo_qa is the complement, not a duplicate: noise applies light damage anywhere and leaves the answer alone, teaching the model to read through it and answer; typo_qa applies heavy damage to one load-bearing word, where guessing is wrong and asking is right. Keeping the two separate is what should stop the model querying every faint typo.

Caveat: the character-level operations mimic OCR artifacts — the exact class of corruption that may have been cleaned out of the source corpus. That is intentional on the question side. If the model begins reproducing OCR-style misspellings in its answers, suspect this first.

Reproduction

Deterministic for a given (count, seed); default seed 1930.

python -m synth.robustness.mine --scan-bytes 0        # mine the clause pools (slow)
python -m synth.robustness.mine --fragments 2500      # sentence pool
python -m synth.robustness.mine --utterances 4000     # utterance pool
python -m synth.robustness.build --dry-run            # counts only
python -m synth.robustness.build                      # build and push

Layout is rows/<route>/part-*.jsonl.