---
pretty_name: Financial Form Normalization Instructions
license: apache-2.0
language:
- en
task_categories:
- text-generation
- text-classification
tags:
- instruction-tuning
- form-normalization
- information-extraction
- financial-forms
size_categories:
- 1K<n<10K
configs:
- config_name: training
  default: true
  data_files:
  - split: train
    path: data/train.jsonl
- config_name: project_diagnostic
  data_files:
  - split: diagnostic
    path: evaluation/project_diagnostic.jsonl
- config_name: release_diagnostic
  data_files:
  - split: diagnostic
    path: evaluation/release_diagnostic.jsonl
---

# Financial Form Normalization Instructions

**3,657 English instruction examples across 38 tasks**, associated with [sraivante/TinyLlama-1.1B-Financial-Form-Normalizer-LoRA](https://huggingface.co/sraivante/TinyLlama-1.1B-Financial-Form-Normalizer-LoRA).

The examples teach short user replies to map to predefined application field values: dates, amounts, ZIP codes, yes/no or boolean values, category labels, navigation intents and small stage/state JSON objects. The intended task is supplied by a system prompt.

**Copyright (c) 2026 sraivante**, for original dataset contributions and original release material, to the extent copyright subsists. Licensed under [Apache License 2.0](LICENSE). See [COPYRIGHT.md](COPYRIGHT.md) for scope and provenance limits.

## Contents and split meaning

| File / configuration | Rows | Role |
|---|---:|---|
| `data/train.jsonl` / `training` | 3,657 | Entire associated training corpus |
| `evaluation/project_diagnostic.jsonl` / `project_diagnostic` | 114 | Existing project examples; all 114 also occur in training |
| `evaluation/release_diagnostic.jsonl` / `release_diagnostic` | 12 | New release-authored diagnostic inputs |
| `evaluation/release_results.json` | 12 paired results | Unchanged base versus adapter, with raw outputs |

There is **no independent validation or test split from training**. The 114 project examples are not held out. The 12 new inputs are absent verbatim from the training corpus, but form a small convenience sample for release checks, not a representative benchmark. Neither diagnostic collection was used to train or tune this published adapter during release preparation.

## Training schema

Every training row has:

- `messages`: exactly three objects in `system`, `user`, `assistant` order, each with string fields `role` and `content`.
- `task`: the task identifier.

```json
{
  "messages": [
    {"role": "system", "content": "Task-specific mapping and output rules"},
    {"role": "user", "content": "Short user reply"},
    {"role": "assistant", "content": "Expected field value"}
  ],
  "task": "task_identifier"
}
```

The diagnostic files use string fields `task`, `system`, `user` and `expected`. The new release diagnostic additionally has a `source` explanation.

The training content is preserved unchanged. There are no exact duplicate rows, but **12 identical system/user pairs have conflicting assistant labels**. For example, some yes/no tasks assign different labels to the same short acknowledgement. Details are recorded in [provenance/data_audit.json](provenance/data_audit.json).

## Tasks

| Task | Training rows |
|---|---:|
| investment_profile | 153 |
| years_to_save | 122 |
| magi_tier | 161 |
| part_a | 96 |
| part_b | 90 |
| part_d | 89 |
| dental | 81 |
| tobacco_usage | 74 |
| supplement | 88 |
| current_savings | 69 |
| pre_medicare_plan | 112 |
| healthcare_profile_state | 73 |
| ltc_profile_state | 77 |
| user_profile_intent | 83 |
| funding_modification_decision | 79 |
| healthcare_funding_decide | 60 |
| ltc_iul_decide | 60 |
| healthcare_state_extraction | 79 |
| level_of_protection | 59 |
| match_type | 55 |
| tranche_size | 152 |
| change_state_identifier | 145 |
| go_to_home_identifier | 54 |
| user_profile_field_detection | 109 |
| change_state_detection | 44 |
| quality_of_care | 70 |
| assisted_care_year | 109 |
| ltc_state_extraction | 61 |
| gender | 82 |
| birthdate | 145 |
| health_profile | 103 |
| life_expectancy | 156 |
| retirement_age | 120 |
| current_zipcode | 163 |
| retirement_zipcode | 168 |
| tax_filing_status | 52 |
| concierge_charge | 54 |
| illustration_choice | 110 |


There are 38 tasks, with 44–168 examples each. Labels and meanings are specific to each system prompt. An integer may represent a field value in one task and a category label in another.

## Load the data

```python
from datasets import load_dataset

repo = "sraivante/Financial-Form-Normalization-Instructions"
train = load_dataset(repo, "training", split="train")
project_checks = load_dataset(repo, "project_diagnostic", split="diagnostic")
release_checks = load_dataset(repo, "release_diagnostic", split="diagnostic")

assert len(train) == 3657
assert len(project_checks) == 114
assert len(release_checks) == 12
```

Pin a repository commit through `revision=...` when reproducing an experiment. `SHA256SUMS` and the provenance files identify the included bytes.

The training notebook renders `messages` through TinyLlama's chat template and performs supervised fine-tuning. The `task` column is metadata, not an additional model input. Re-tokenization with the saved tokenizer produced 64–302 tokens per training conversation.

## Provenance and historical association

The source is `fp_finetune_jsonl/chat/all_combined.jsonl`, a local project corpus dated May 20, 2026. The 38 individual task files reproduce the combined messages in the same order once their task identifiers are attached.

A second copy, `LLM_FINE_TUNE/finetune/data/datasets/all_combined-uploaded-20260520-125234.jsonl`, has identical parsed records and order, with different serialization bytes. The TinyLlama notebook loads `all_combined.jsonl` and records **3,657 examples**, matching this corpus. Its training history matches the adapter's 1,374-step, three-epoch checkpoint.

No checksum of the original Colab input file was recorded. The available evidence strongly supports this corpus's association, but does not prove historical byte identity. Full source hashes and the precise association limits are in [provenance/source.json](provenance/source.json).

The corpus consists of short, example-style field values, phrase variants and misspellings. No generator script or external dataset attribution accompanies this exact chat corpus; the proportion authored manually versus with model assistance is not established. A separate later Qwen generation pipeline is not treated as evidence of how this snapshot was created.

Publication preserved the original training and project diagnostic files without editing, label correction, filtering, shuffling or deduplication. The 12 release diagnostic cases and their results were added separately and clearly labeled.

## Evaluation and limitations

| Model | Exact requested values |
|---|---:|
| TinyLlama chat base | 0 / 12 |
| Published LoRA adapter | 11 / 12 |

The adapter returned RETRY for the jointly filing tax-status example, whose expected
label was MARRIED_FILING_JOINTLY. The base generally returned explanations instead
of the required bare values. Only surrounding whitespace was removed for scoring;
this measures strict value-and-format matching on these twelve inputs.


Scoring strips surrounding whitespace and requires an exact target string. Explanations count as failures even if they contain the correct value. The comparison uses the same unquantized BF16 base and task prompts, with greedy generation capped at 32 new tokens. Full conditions and outputs are in [evaluation/release_results.json](evaluation/release_results.json).

The task rules include application-specific defaults and fixed mappings, such as a default retirement ZIP and categorical numeric values. They are historical demonstration rules, not validated current tax, Medicare, medical, investment-return or eligibility guidance. Applying those values in a real product requires independent validation of the product's rules and the user's intended value.

The dataset includes example dates, ZIP codes, health-related labels and other form fields. No external customer-record source is documented. Pattern checks found no email addresses, credential-like strings, phone-number patterns or SSN-format strings in the training corpus; these checks do not establish that every value is synthetic or comprehensively de-identified.

Known constraints include conflicting labels, narrow gender categories, fixed defaults, small task-specific sample counts, no original held-out split, and limited coverage of ambiguous or novel replies. English is the intended language.

## Copyright and license

Original contributions are offered under Apache License 2.0 with **Copyright (c) 2026 sraivante**, subject to the rights the contributor holds. No ownership is claimed over third-party models or tokenizer assets. The dataset does not redistribute the TinyLlama base weights.

Related model: [TinyLlama 1.1B Financial Form Normalizer LoRA](https://huggingface.co/sraivante/TinyLlama-1.1B-Financial-Form-Normalizer-LoRA). A matching dataset snapshot is also included under that model repository's `training_data/` directory.

Release version: **2026-09-25**.