id string | messages list | notice_id string | genre string | intents list | provenance dict | contrast_group string | contrast_role string | contrast_term string | v4_source string | v4_softened_modal_advisories int64 |
|---|---|---|---|---|---|---|---|---|---|---|
train-reduction-0000-a | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-reduction-0000 | reduction | [
"open",
"contrast_term",
"confirmation",
"advice_bait"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 3
} | cg-0000 | term_used | allotment | v3-survivor | 0 |
train-reduction-0000-b | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-reduction-0000 | reduction | [
"open",
"contrast_term",
"emotional"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 2
} | cg-0000 | term_absent | allotment | v3-survivor | 1 |
train-termination-0001-a | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-termination-0001 | termination | [
"open",
"contrast_term"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 0
} | cg-0001 | term_used | recertification | v3-survivor | 0 |
train-reduction-0004-b | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-reduction-0004 | reduction | [
"open",
"contrast_term"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 2
} | cg-0004 | term_absent | redetermination | v3-survivor | 1 |
train-termination-0005-b | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-termination-0005 | termination | [
"open",
"contrast_term",
"deadline_collapse"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 1
} | cg-0005 | term_absent | redetermination | v3-survivor | 1 |
train-recert_form-0006-a | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-recert_form-0006 | recert_form | [
"open",
"contrast_term"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 0
} | cg-0006 | term_used | recertification | v3-survivor | 1 |
train-verification-0007-a | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-verification-0007 | verification | [
"open",
"contrast_term"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 0
} | cg-0007 | term_used | verification | v3-survivor | 1 |
train-reduction-0008-a | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-reduction-0008 | reduction | [
"open",
"contrast_term",
"emotional",
"vocab"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 2
} | cg-0008 | term_used | redetermination | v3-survivor | 1 |
train-reduction-0008-b | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-reduction-0008 | reduction | [
"open",
"contrast_term",
"emotional",
"advice_bait"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 2
} | cg-0008 | term_absent | redetermination | v3-survivor | 0 |
train-termination-0009-b | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-termination-0009 | termination | [
"open",
"contrast_term"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 0
} | cg-0009 | term_absent | recertification | v3-survivor | 0 |
train-reduction-0016 | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-reduction-0016 | reduction | [
"open",
"emotional",
"confirmation",
"advice_bait"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 3
} | null | null | null | v3-survivor | 1 |
train-verification-0020 | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-verification-0020 | verification | [
"open",
"vocab",
"form_field"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 1
} | null | null | null | v3-survivor | 1 |
train-reduction-0023 | [
{
"role": "system",
"content": "You are an assistant that explains benefits notices — letters about SNAP,\nMedicaid, SSI, and similar programs — to the adult who received them. Many\nreaders have intermediate English (including second-language readers). You\nexplain the letter the user gives you. You never ... | train-reduction-0023 | reduction | [
"open",
"emotional"
] | {
"teacher": "claude-sonnet-5",
"date": "2026-08-18T00:00:00",
"repairs": 1
} | null | null | null | v3-survivor | 1 |
End of preview. Expand in Data Studio
Benefits Notices, Explained — training set v4
The dataset is the deliverable. 124 checker-filtered teacher-distillation dialogs that train a small model (Qwen3-4B QLoRA) to explain U.S. benefits notices under a falsifiable behavior spec: earned-vocabulary ceiling (frozen 2,801-lemma NGSL allowed list + words the reader used + glossed terms), character-for-character anchor fidelity (dates, amounts, phones, durations, case/form numbers, citations), quote-then-explain, no advice, mandated deadline banner on adverse actions.
- Format:
train.jsonl, chat format (messageswith system/user/ assistant), one dialog per line, 2–5 assistant turns each. - Filter: every assistant turn passed the deterministic compliance checker (spec v3, checker v1.2 — rules, calibration audit, and FP gate in the project repo). One repair round per failing turn; second failure discards the whole dialog.
- Teacher: claude-sonnet-5, conditioned on the behavior spec (stored as each dialog's system message). v4 additionally used a generation-only fidelity addendum, an obligation-softening repair round, and a dialog-level advisory gate (≤1) — the stored system message is the plain deploy prompt, so eval deltas attribute to data, not prompts.
- v4 delta over v3 (the measured point of this version): obligation-
softening density cut 1.60 → 0.47 advisories/dialog (3.4×) by
dropping the 54 worst v3 dialogs and regenerating under the pressures
above; the model trained on v4 softened "must" 33 → 7 times on the same
frozen eval (4.7×). Full provenance:
DATASET.md,STATS.json,DROPS.json. - Letters are synthetic (template + programmatic variation, no LLM, no
real PII:
A. Sample,555-01xx,00-…-SNcase numbers) and are hard-checked disjoint from the project's evaluation set.
Trained checkpoint: jmerithew1/qwen3-4b-benefits-notice-qlora.
Generation/filter code, checker rules, calibration audits, and eval
harness: labs.gauntletai.com/jamesmerithew/slm-week7 (GauntletAI Week 7).
- Downloads last month
- 10