v1: sft, rl_pool, dev

README.md

@@ -0,0 +1,61 @@
+---
+license: cc-by-4.0
+language: [en, hi]
+tags: [math, finance, india, hinglish, reasoning, verifiable-rewards, grpo, synthetic]
+pretty_name: IndicFinance Reasoner data (v1)
+size_categories: [1K<n<10K]
+configs:
+- config_name: sft
+  data_files:
+  - split: train
+    path: data/sft.jsonl
+- config_name: rl_pool
+  data_files:
+  - split: train
+    path: data/rl_pool.jsonl
+- config_name: dev
+  data_files:
+  - split: validation
+    path: data/dev.jsonl
+---
+
+# IndicFinance Reasoner data (v1)
+
+Indian personal-finance calculations (GST, discounts, simple/compound interest, EMI) as step-by-step word problems in
+**English, Hindi (Devanagari) and Hinglish (romanised)**, for training and evaluating verifiable reasoning.
+Built for the Lexsi AI Safety & Alignment Hackathon (Track 1, PS 6). Code: https://github.com/tanmay-srivastav4/indic_finance
+
+## How it is made
+- Generated by deterministic code (seed 1234): 7 training families, 4-5 phrasings per family per language,
+  Indian number formats (lakh/crore, Rs./INR/₹, some Devanagari digits).
+- Every answer **and every intermediate step value** is computed exactly (Python `Decimal`) and every solution is
+  re-checked by the verifier before it is written: correct by construction.
+- ~15% of items are **unanswerable**: one required number is removed; the correct answer is `INSUFFICIENT`.
+- Splits are assigned by `group_id` (the 3 language renderings of one problem share it), so no problem crosses splits.
+
+## Configs
+| Config | Rows | Use |
+|---|---|---|
+| `sft` | ~3,000 | supervised fine-tuning targets (code-written solutions) |
+| `rl_pool` | 3,765 | problems for GRPO / rejection sampling |
+| `dev` | 150 | development checks (7 families x 3 languages; 27 unanswerable) |
+
+Test splits (`test_infamily` 450, `test_heldout` 150 with 2 unseen families, `test_abstain` 102) will be added after
+the final evaluation.
+
+## Answer format
+```
+<numbered steps, in the question's language>
+calc: <one arithmetic expression using the question's numbers>
+\boxed{<final answer>}        # \boxed{INSUFFICIENT} when a required number is missing
+```
+Correct = exactly one box, value within the family tolerance (`target.tol_abs` / `target.tol_rel`).
+
+## Fields
+`id`, `group_id`, `task_family`, `tier` (1-3), `language`, `template_id`, `params`, `question`, `answerable`,
+`target {type, value, tol_abs, tol_rel}`, `reference_calc`, `steps [{kind, expr, value}]`, `solution`,
+`question_numbers`, `missing_param`, `split`, `seed`, `generator_version`.
+
+## Limitations
+Template-generated: wording is less varied than real user questions, so in-distribution scores are optimistic.
+Hindi/Hinglish templates were written by the team and not yet natively reviewed. Not financial advice.