spkc83/retail-bank-servicing-alignment-sft

Retail Bank Servicing Alignment SFT

The training corpus for the Granite retail-bank servicing agent. It is the released tool-use SFT corpus merged with a servicing-alignment continuation curriculum that teaches multi-turn behaviours the base corpus does not: what to do when the customer says "that one", when a policy question interrupts a transfer, when the agent's own previous turn was wrong, and when the honest answer is that the agent cannot see what it was asked about.

Every conversation is self-authored synthetic data about a fictional bank. It contains no real customers, credentials, accounts, or financial events.

Splits

Split	Rows	Alignment rows	Base rows	Allowed use
train	3,959	3,118	841	granite-continuation-sft
validation	447	268	179	granite-continuation-evaluation
test	215	35	180	granite-continuation-evaluation

Three further files are not trainable and are declared as such in manifest.json:

File	Rows	Purpose
coreference-shadow.jsonl	32 (16 pairs)	Post-selection behavioural gate; never used for checkpoint selection
granite-v7-shadow.jsonl	13	Checkpoint selection and generalization evaluation
screenshot-regression.jsonl	9	Regression fixture for the demo surface

manifest.json is the authority on what each file may be used for: every entry carries an allowed_use list and the non-trainable files carry trainable: false. Read it before wiring the corpus into a trainer rather than inferring intent from filenames.

What the alignment curriculum teaches

Counts are train rows. The largest families are the deictic-coreference pairs, where the same history is followed by a turn that either resolves to one card or is genuinely ambiguous, and the model must act in the first case and ask in the second.

Family	Rows	Behaviour
deictic_replace_action / _ambiguity	784 / 784	Act on a resolved referent; ask when two candidates fit
long_context_tool_fidelity	200	Keep the tool contract when a long history ends in a misleading turn
history_entity_action / _ambiguity	128 / 32	Resolve an entity named earlier in the conversation
no_evidence_honesty	112	Do not claim a fact the tools did not return
capability_boundary, credential_hygiene, scope_refusal	84 each	Decline what the agent cannot or must not do
deictic_*_clarification	72 each	Clarify when the referent is ineligible or missing
agent_repair	64	Recover when the agent's own previous turn was wrong
card_anaphora_action, clarification_answer, service_case_context	64 each	Carry a referent across turns
deictic_replace_reinforcement_*	64 each	The same mapping over a wider entity pool
banking_topic_shift, external_topic_shift, policy_detour, policy_resume	32 each	Leave and return to a servicing task
natural_social_style	12	Answer a greeting without starting a transaction

Provenance

Rows are generated deterministically, then two teacher passes rewrite wording only:

Pass	Rows moved	Field
Finals realization	651 train + 168 validation	final_response
Prompt realization	366 train + validation	user_content

Both are safe by construction. An immutable record hash covers the tool calls, the tool results, the grounding facts, and the split keys, and deliberately excludes the two wording fields — so a teacher can rephrase a question or an answer but can never edit what makes a row's supervision correct. A pass that tried would fail the hash check at build time. Each pass records the model it credits in every touched row's provenance.

The prompt pass exists because the evaluation splits had become the training questions with a different trailing phrase. Rows whose prompt moved are held to an isolation invariant: a rewritten prompt may not share a 4-gram with any test, shadow, or regression prompt.

Build-time gates

Generation fails rather than reports. The build asserts, among others:

• PII: zero matches for emails, SSNs, and long digit runs across every message, not only the trainable pair (pii_matches: 0).
• Held-out isolation: no test-split current turn appears in train, and no long n-gram from a held-out row leaks into train (both empty).
• Trainable-text word ban: train and validation may not say app, demo, synthetic, sandbox, test, and similar — a model should not learn to talk about a product surface it cannot see. The frozen evaluation splits are exempt by split so their fixtures stay byte-identical.
• Coreference pairing: a normalized current turn may group multiple pairs only across distinct history forms, so a pair cannot be won by memorising the question.

Contamination

Measured two ways, because the obvious one misleads on a domain corpus:

Split	Median nearest-train similarity	≥0.95	≥0.90
test	0.843	0	8
validation	0.789	0	34

Raw 4-gram overlap over-reports badly here: a banking corpus shares its own vocabulary ("my card ending in", "freeze the active card"), and counting that as contamination would argue for training a model to avoid the nouns of its own domain. Nearest-neighbour similarity separates the cases — a test question that differs from a training question only in a trailing phrase scores above 0.95, and no row now does.

Reproducing it

The corpus is generated, not collected. See the generator — scripts/retail_bank/prepare_servicing_alignment_data.py with the two teacher files reproduces this revision byte for byte, and manifest.json carries the SHA-256 of every file plus the digests of the base corpus, the synthetic bank, and the policy corpus it was built from.

Intended use and limits

This is a research demonstration of governed synthetic SFT data, not a banking product. The behaviours it teaches are enforced at the data layer and gated at training time; nothing here is a substitute for the safety architecture of the serving harness, which carries the guarantees this corpus can only bias toward.