Ling-Tiny-SFT

An SFT dataset for structured output in composite payment workflows.

Ling-Tiny-SFT turns orders, payment channel balances, and promotion rules into payment-request-to-JSON examples for Ling Tiny model fine-tuning. The model must calculate the payment plan for the selected channel and recommend the feasible channel offering the largest discount, while handling split payments, exclusive promotions, rounding, and error conditions.

Prepared for the AReno Ling Tiny end-to-end fine-tuning recipe, the dataset contains 1,400 training examples and 100 validation examples. Requests are generated programmatically, and answers are computed by a reference payment service. Please note the dataset contains no real transaction records, all training targets are direct JSON outputs without <think> tags or reasoning text.

Dataset Size and Scenario Distribution

Split	File	Examples	Purpose
train	train.jsonl	1,400	SFT parameter updates
validation	valid.jsonl	100	Configuration or checkpoint selection during development
Total	—	1,500	—

This version does not include a test split. The validation set does not replace an independent final evaluation set.

Scenario	Original Label	train	validation
Split payment	SPLIT_PAYMENT	357	25
Recommended channel switch	RECOMMEND_SWITCH	264	19
Exclusive vs. stackable promotions	EXCLUSIVE_VS_STACK	213	15
Rounding boundary	ROUND_HALF_UP_BOUNDARY	203	15
Single-channel payment	SINGLE_CHANNEL_NORMAL	148	11
Recommendation tie	TIE_BREAKER	140	10
Insufficient funds	INSUFFICIENT_FUNDS	45	3
Selected channel not found	CHANNEL_NOT_FOUND	30	2

Scenario labels reflect the original generator's sampling intent. They are not independently revalidated classifications of the boundary conditions in each example.

Repository Structure

Ling-Tiny-SFT/
├── README.md             # Dataset overview, fields, and loading instructions
├── train.jsonl           # 1,400 training examples
├── valid.jsonl           # 100 examples mapped to the validation split
├── system_prompt.txt     # Shared payment rules, already included in each prompt
└── manifest.json         # Provenance, preparation rules, statistics, and checksums

Data Format

Each line in a JSONL file contains one example:

Field	Type	Description
case_id	string	Original example identifier
source_line	integer	One-based line number in the source file
bucket	string	Original scenario label
request_fingerprint	string	Request SHA-256 excluding the order ID
input	object	Order, channels, promotions, and selected channel
ground_truth	object	Expected result from the reference payment service
prompt	string	Complete rules and compact request JSON, used as model input
response	string	Expected result serialized as JSON, used as the supervised target

Example from the Training Set

The following JSON content shows case_id, the complete input, and ground_truth for training example SFT_001262; other fields are omitted for readability. The order amount is 7 cents. The selected channel receives a 1-cent general discount and a 1-cent channel discount, resulting in a 5-cent deduction. The same channel is also recommended.

{
  "case_id": "SFT_001262",
  "input": {
    "available_channels": [
      {
        "balance": 7,
        "channel_id": "GIFT_VOUCHER_CARD_9414"
      },
      {
        "balance": 21,
        "channel_id": "BANK_CMB_GOLD_6278"
      },
      {
        "balance": 29553,
        "channel_id": "BANK_ICBC_DEBIT_7379"
      },
      {
        "balance": 9,
        "channel_id": "BANK_SPDB_CARD_9218"
      }
    ],
    "order": {
      "order_amount": 7,
      "order_id": "ORD_0011261"
    },
    "promotions": [
      {
        "applicable_channel_id": null,
        "discount_amount": 1,
        "discount_percent": 0,
        "is_exclusive": false,
        "promotion_id": "PROMO_FIXED_1",
        "promotion_type": "FIXED_DISCOUNT",
        "threshold": 3
      },
      {
        "applicable_channel_id": "BANK_CMB_GOLD_6278",
        "discount_amount": 1,
        "discount_percent": 0,
        "is_exclusive": false,
        "promotion_id": "PROMO_CH_2",
        "promotion_type": "CHANNEL_DISCOUNT",
        "threshold": 2
      }
    ],
    "selected_channel_id": "BANK_CMB_GOLD_6278"
  },
  "ground_truth": {
    "cashier_receipt": {
      "applied_promotions": [
        "PROMO_FIXED_1",
        "PROMO_CH_2"
      ],
      "payable_amount": 5,
      "primary_channel_id": "BANK_CMB_GOLD_6278",
      "primary_deduct_amount": 5,
      "secondary_channel_id": null,
      "secondary_deduct_amount": 0,
      "total_discount_amount": 2
    },
    "error_message": null,
    "order_amount": 7,
    "order_id": "ORD_0011261",
    "recommended_option": {
      "applied_promotions": [
        "PROMO_FIXED_1",
        "PROMO_CH_2"
      ],
      "payable_amount": 5,
      "primary_deduct_amount": 5,
      "recommended_channel_id": "BANK_CMB_GOLD_6278",
      "secondary_channel_id": null,
      "secondary_deduct_amount": 0,
      "total_discount_amount": 2
    },
    "status": "SUCCESS"
  }
}

Just use prompt and response for training. For inference, pass only prompt: do not append the rules again or include response or ground_truth in the input.

Quick Start

The release is intended to provide identical files through two dataset hubs:

• Hugging Face: the mirror repository ID is pending. Replace AsystemAReno/Ling-Tiny-SFT in the examples below with the published ID.
• ModelScope (AsystemAReno/Ling-Tiny-SFT): the repository has been created; the prepared data files are awaiting upload.

After publication, either option below provides both train.jsonl and valid.jsonl. No additional assets from the other platform are required.

Option 1: HuggingFace

First install the huggingface dependencies, then download the dataset repository:

pip install datasets huggingface_hub

hf download AsystemAReno/Ling-Tiny-SFT \
  --repo-type dataset \
  --local-dir Ling-Tiny-SFT

To load the splits directly in Python instead:

from datasets import load_dataset

dataset = load_dataset("AsystemAReno/Ling-Tiny-SFT")
train = dataset["train"]             # 1,400 examples
validation = dataset["validation"]   # 100 examples

print(train[0])

The CLI command creates the local directory used in the Usage example below. The Python alternative returns the splits directly and uses the library's cache.

Option 2: ModelScope

Install the ModelScope SDK and download the complete dataset repository:

pip install modelscope

modelscope download --dataset AsystemAReno/Ling-Tiny-SFT \
  --local-dir Ling-Tiny-SFT

Alternatively, clone the repository with Git LFS installed:

git lfs install
git clone https://modelscope.cn/datasets/AsystemAReno/Ling-Tiny-SFT.git

Usage

After downloading or cloning the repository, run this example from its parent directory to read both JSONL files and inspect a training example:

import json
from pathlib import Path

data_dir = Path("Ling-Tiny-SFT")

with (data_dir / "train.jsonl").open(encoding="utf-8") as stream:
    train = [json.loads(line) for line in stream if line.strip()]

with (data_dir / "valid.jsonl").open(encoding="utf-8") as stream:
    validation = [json.loads(line) for line in stream if line.strip()]

print(f"Training examples: {len(train)}")
print(f"Validation examples: {len(validation)}")

sample = train[0]
print("Case ID:", sample["case_id"])
print("Scenario:", sample["bucket"])
print("Model input:", sample["prompt"])
print("Training target:", sample["response"])
print("Expected result:")
print(json.dumps(sample["ground_truth"], indent=2, ensure_ascii=False))

Use prompt as the model input and response as the SFT target. For validation, generate a response from prompt and compare the parsed result with ground_truth. This release has no test split; use a separate evaluation set for final reporting.

For reproducible experiments, pin the published revision using --revision in the download commands or revision= in load_dataset. Remote downloads remain unverified until publication; local JSON loading has been verified.

Applications

• Domain SFT: train small models to produce business JSON according to explicit payment rules.
• Structured output: examine field completeness, numeric types, and success or error responses.
• Rule execution analysis: analyze errors in arithmetic, promotion selection, two-channel deductions, and recommendations.
• Reproducible experiments: train on fixed dataset versions and compare results before and after fine-tuning using an independent evaluation set.

AReno based Recipe

This dataset is intended for end-to-end Ling Tiny SFT experiments with AReno. After loading the data, the planned workflow covers baseline evaluation, fine-tuning, adapter saving, reloading in a separate process, and evaluation after fine-tuning. Mac, Colab, and Kaggle use the same fixed data; training compatibility must still be verified separately on each platform.

Target-tokenizer length budgets, chat templates, and thinking-mode settings have not yet been validated. This release makes no claims about model performance or successful training across platforms. Changes to prompts, examples, or splits should produce a new dataset version, rather than silent filtering in the Notebook.