CRFD-25: Railway Food & Drink Dataset

• CRFD-25 is a railway-oriented takeout dataset defined by over 6600 city-level dish entries curated for recommendation alignment in onboard catering services.
• It is constructed from leading food delivery platforms, covering both regional specialties and popular fast food to support personalized dining across diverse age groups.
• The dataset underpins LLM-based conversational recommenders by mapping unconstrained suggestions to valid menu items through feature similarity and operational grounding.

Searching arXiv for the primary CRFD-25 paper and closely related food-recognition/domain-LLM references. arXiv search query: (Li et al., 31 Jul 2025) OR LLM4Rail OR "Chinese Railway Food and Drink" OR ChineseFoodNet OR FoodSky OR "Feature-Enhanced TResNet"

Chinese Railway Food and Drink (CRFD-25) is a publicly accessible takeout dataset tailored for railway services and introduced as the core data asset of the onboard catering module in "LLM4Rail: An LLM-Augmented Railway Service Consulting Platform" (Li et al., 31 Jul 2025). It is presented as the first publicly available railway-oriented takeout dataset, created for a setting in which passengers order meals before departure and restaurants deliver them to the train. Within that system, CRFD-25 is not a generic food corpus: it is the structured item universe that grounds onboard dining recommendation in a realistic railway service inventory, supports customized dining services onboard trains, and constrains open-ended LLM outputs to recommendations that are operationally usable (Li et al., 31 Jul 2025).

1. Origin, purpose, and railway-service context

CRFD-25 was created in response to a practical service problem rather than as a standalone academic benchmark. The motivating scenario is that railway operators seek value-added, personalized services beyond basic ticketing, and the paper identifies food ordering as a particularly promising opportunity. The envisioned workflow is that passengers place food orders online before boarding, restaurants prepare the food, and the food is delivered to the train (Li et al., 31 Jul 2025).

Within that setting, a generic LLM is insufficient because it may recommend foods that are unavailable, irrelevant to the departure city, or nonexistent in the actual railway catering inventory. CRFD-25 therefore functions as a grounded menu space for railway service consulting. The paper repeatedly frames it as enabling customized dining services onboard trains and as the item space over which an LLM-based conversational recommender can operate safely (Li et al., 31 Jul 2025).

The dataset is described as covering a wide range of signature dishes categorized by cities, cuisines, age groups, and spiciness levels. This emphasis is central to its definition. City context matters because railway catering is geographically situated; cuisine matters because route-relevant and region-relevant specialties are part of the service logic; age groups and spiciness levels support personalization rather than generic retrieval. The paper does not explain the meaning of the “25” in CRFD-25, so its interpretation should not be inferred (Li et al., 31 Jul 2025).

2. Construction, geographic coverage, and curation principles

CRFD-25 was collected from Meituan and Dianping, identified as two of the leading online food delivery platforms in China. Its geographic scope consists of 33 high railway traffic cities. In each city, more than 200 signature dishes are manually picked and labeled from different restaurants (Li et al., 31 Jul 2025).

The paper states two explicit inclusion principles. First, the dataset maintains a balance between regional specialties and widely consumed fast food items. Second, it is designed to cater to a wide range of age groups. These principles are illustrated by concrete examples. In Shanghai, the dataset includes both Shanghai Soup Dumplings as a local specialty and McDonald’s as globally popular fast food. It also includes hamburgers, described as favored among children, and Mango Pomelo Sago, described as especially popular among women and young people (Li et al., 31 Jul 2025).

From the explicit counts, the guaranteed lower bound is $33 \times 200 = 6600$ city-level dish entries. The paper does not explicitly state the final number of unique records, whether some items are duplicated across cities, or what the post-cleaning total count is. It also does not specify whether beverages are included at the same scale as food, although the system formulation refers to food and beverage item set $\mathcal{I}$ (Li et al., 31 Jul 2025).

The curation process is clearly manual, but its methodological transparency is partial. The paper states that dishes were manually picked and labeled, and Figure 1 illustrates feature annotation. However, it does not provide annotator count, annotation guidelines, inter-annotator agreement, quality-control protocol, labeling tool/software, exact cleaning pipeline, exact exclusion criteria, deduplication methods, or data normalization procedures. The effective curation rules are nevertheless clear: select representative or signature dishes, preserve a mix of local specialties and common fast food, include foods suitable for different age groups, and attach structured labels for recommendation (Li et al., 31 Jul 2025).

3. Data representation and documented schema

Each CRFD-25 item is stored in the form of an illustrative image along with its relevant features. The paper and figure description directly document several fields (Li et al., 31 Jul 2025).

Field	Documented content	Encoding/status
City	33 railway-relevant cities	Explicitly described
Dish/item name	Examples include Shanghai Soup Dumplings, Chongqing Noodle, Mango Pomelo Sago	Explicitly implied
Type of food	1 = Chinese food; 2 = Western food	Sequential integer encoding
Cuisine	1 = Sichuan cuisine; 2 = Cantonese cuisine; 3 = Shandong cuisine; etc.	Sequential integer encoding
Meal suitability	Examples include breakfast and lunch	Explicitly described
Child-friendliness / age groups	Personalization signal tied to age groups	Explicitly described, exact encoding not formalized
Spiciness level	not spicy, mild, medium, very spicy, extra spicy	Multi-hot encoding
Image	Illustrative image for each item	Explicitly described

The paper explicitly gives two encoding patterns. Some categorical fields use sequential integer encoding, as illustrated by type of food and cuisine. Multi-select features use multi-hot encoding. The spiciness example is the most detailed: for Chongqing Noodle, not spicy, mild, medium, very spicy, and extra spicy are represented using multi-hot encoding, with the latter four marked 1 and not spicy marked 0 in the given example (Li et al., 31 Jul 2025).

The schema is only partially formalized. The paper does not provide a full schema table, and it does not clearly state whether CRFD-25 includes restaurant name, price, portion size, nutrition/allergens, stock/availability, station coverage, packaging suitability, delivery lead time, or beverage subtype labels. The introduction once mentions “price ranges, and other relevant features,” but the later dataset description does not define price as a field. A cautious reading is that price is suggested in prose but not concretely documented in the available text (Li et al., 31 Jul 2025).

This representation design is important because the dataset is not merely an image collection. Its operational role depends on structured metadata. City, cuisine, meal suitability, age-group suitability, and spiciness are the attributes that enable recommendation alignment rather than unrestricted text generation.

4. Formal role in LLM4Rail and conversational recommendation

Within LLM4Rail, CRFD-25 is the item universe $\mathcal{I}$ for the railway catering recommender. The paper formalizes the interaction using passenger set $\mathcal{P}$, utterance set $\mathcal{U}$, and food and beverage item set $\mathcal{I}$. A passenger $p_t \in \mathcal{P}$ produces utterance $u_t \in \mathcal{U}$ at turn $t$, and the conversation is represented as

$$\mathcal{C}_{1:T} = (p_t, u_t, \mathcal{I}_t)_{t=1}^T,$$

where $\mathcal{I}$0 is the set of items mentioned in turn $\mathcal{I}$1, possibly empty (Li et al., 31 Jul 2025).

The recommendation module is an LLM-based zero-shot conversational recommender. At turn $\mathcal{I}$2, it feeds dialogue history $\mathcal{I}$3, task description $\mathcal{I}$4, and format instruction $\mathcal{I}$5 to the LLM, producing top-$\mathcal{I}$6 preliminary recommendations: $\mathcal{I}$7 The paper states that $\mathcal{I}$8 prompts recommendation based on dialogue context and profiles such as gender, age, and the place of birth, while $\mathcal{I}$9 constrains output format. The exact prompt text is not provided (Li et al., 31 Jul 2025).

CRFD-25 plays two roles in this architecture. First, it is the knowledge base or candidate inventory: it defines what dishes can actually be recommended in railway service. Second, it is the feature space for alignment: its structured labels allow free-form LLM suggestions to be mapped to valid items through feature similarity (Li et al., 31 Jul 2025).

This second role is algorithmically central because unconstrained LLM outputs need not belong to the valid inventory: $\mathcal{I}$0 The system therefore applies a mapping function

$\mathcal{I}$1

where $\mathcal{I}$2 is based on feature similarity between generated items and CRFD-25 items. The paper refers to this as recommended item alignment (Li et al., 31 Jul 2025).

A concrete example clarifies the mechanism. If the LLM recommends Angus Beef Burger, which is not present in CRFD-25, the system may instead recommend Spicy Chicken Burger as the aligned alternative. This shows that the mapping is intended to preserve feature-level semantic similarity rather than rely on exact string identity (Li et al., 31 Jul 2025).

The food recommender is embedded in the broader Question-Thought-Action-Observation (QTAO) prompting framework. At iteration $\mathcal{I}$3 of turn $\mathcal{I}$4,

$\mathcal{I}$5

and the next thought and action are generated by

$\mathcal{I}$6

Algorithm 1 lists Food & Drink Recommendation as one action among Ticketing, Weather, and ChitChat. CRFD-25 is therefore not an isolated dataset; it is a tool-facing substrate inside an iterative LLM agent loop (Li et al., 31 Jul 2025).

5. Empirical behavior, alignment effects, and operational significance

The paper evaluates the catering module with Qwen3-235b-a22b, GPT-4o, and Gemini-2.5-pro. Because there is a scarcity of food-and-drink recommendation dialogues, the evaluation uses a user simulator to interact with the conversational recommenders (Li et al., 31 Jul 2025).

Two metric families are reported. The first is $\mathcal{I}$7, the proportion of top-$\mathcal{I}$8 generated recommendations that appear in CRFD-25. The reported values are:

• Qwen3: $\mathcal{I}$9, $\mathcal{P}$0, $\mathcal{P}$1
• GPT-4o: $\mathcal{P}$2, $\mathcal{P}$3, $\mathcal{P}$4
• Gemini-2.5-pro: $\mathcal{P}$5, $\mathcal{P}$6, $\mathcal{P}$7

For all tested models and all reported $\mathcal{P}$8, less than 60% of recommendations are present in CRFD-25. This is the empirical justification for the alignment step: open-ended recommendation quality in language does not imply inventory validity (Li et al., 31 Jul 2025).

The second metric family is $\mathcal{P}$9, reported for zero-shot recommendation with and without alignment. For Qwen3, zero-shot yields $\mathcal{U}$0, $\mathcal{U}$1, and $\mathcal{U}$2, while zero-shot w/ Alignment yields $\mathcal{U}$3, $\mathcal{U}$4, and $\mathcal{U}$5. For GPT-4o, zero-shot yields $\mathcal{U}$6, $\mathcal{U}$7, and $\mathcal{U}$8, while zero-shot w/ Alignment yields $\mathcal{U}$9, $\mathcal{I}$0, and $\mathcal{I}$1 (Li et al., 31 Jul 2025).

The paper interprets these results as evidence that feature-matching alignment is effective. For Qwen3, it reports gains of $\mathcal{I}$2, $\mathcal{I}$3, and $\mathcal{I}$4 at $\mathcal{I}$5. For GPT-4o, it reports gains of $\mathcal{I}$6 and $\mathcal{I}$7 at $\mathcal{I}$8. An important nuance is that GPT-4o’s $\mathcal{I}$9 drops from 63.20% to 54.40% after alignment. The improvement claim therefore applies to $p_t \in \mathcal{P}$0 and $p_t \in \mathcal{P}$1, not to $p_t \in \mathcal{P}$2 (Li et al., 31 Jul 2025).

These experiments define CRFD-25’s operational significance. The dataset is not used merely for corpus release or offline annotation. It acts as an active constraint and feature substrate that converts unconstrained LLM recommendations into deployable railway-catering outputs.

6. Methodological position, limitations, and adjacent research

CRFD-25 occupies an unusual position among food datasets. It is railway-oriented, recommendation-oriented, and feature-structured, rather than primarily an image-classification benchmark. That distinction matters when situating it relative to adjacent arXiv work.

ChineseFoodNet is a large-scale image dataset for Chinese dish recognition, containing 185,628 images across 208 categories and framed as single-label image classification (Chen et al., 2017). Its documented emphasis on mixed web imagery and real-life photos, along with semi-automatic cleaning and labeling, is methodologically relevant to CRFD-25. However, CRFD-25 is introduced for recommendation grounding and railway service consulting rather than for dish recognition as a primary task.

"Feature-Enhanced TResNet for Fine-Grained Food Image Classification" reports that FE-TResNet improves classification accuracy to 81.37% on ChineseFoodNet and 80.29% on CNFOOD-241 by adding StyleRM and DCA to TResNet (Liu et al., 17 Jul 2025). CRFD-25 stores each item with an illustrative image along with its relevant features, so this suggests a natural methodological interface with fine-grained food recognition. A plausible implication is that subtle texture, packaging, garnish, and presentation cues could matter if CRFD-25 is extended toward image-based railway food/drink classification. The paper introducing CRFD-25, however, does not present such a benchmark.

FoodSky is a food-oriented LLM built on FoodEarth, a Chinese food-domain instruction corpus of 811,491 question-answer pairs, and evaluated on chef and dietetic examinations (Zhou et al., 2024). Its emphasis on retrieval-augmented food reasoning, dietary science, flavor profiles, food safety measures, food recipes, and healthy eating principles suggests a broader food-intelligence layer that is conceptually adjacent to CRFD-25. Within LLM4Rail, though, the CRFD-25 contribution is more specific: inventory grounding for railway catering recommendation rather than general culinary reasoning.

Several limitations of CRFD-25 are explicit. The paper does not provide a full schema table, exact cleaning/filtering algorithm, exact feature similarity function $p_t \in \mathcal{P}$3, or exact counts for total unique items, number of cuisines, number of age-group labels, number of restaurants, or number of images (Li et al., 31 Jul 2025). It also notes a practical tension: although the corpus could be expanded, only a limited number of items can be recommended due to spatial and temporal constraints. The strongest empirical limitation is that $p_t \in \mathcal{P}$4 across tested LLMs, indicating that raw model knowledge substantially exceeds the curated inventory (Li et al., 31 Jul 2025).

Other constraints are more interpretive. Because dishes were manually picked and labeled, a plausible implication is that the dataset reflects curator judgments about what counts as a signature dish, which foods are age-appropriate, and what balance should be maintained between regional specialties and mainstream items. The examples associating hamburgers with children and Mango Pomelo Sago with women and young people are explicitly given in the paper; a plausible implication is that demographic personalization relies partly on culturally contingent heuristics. The paper does not discuss fairness or bias in detail (Li et al., 31 Jul 2025).

Taken together, these features define CRFD-25 as a grounded railway catering dataset whose principal importance lies in operational alignment. Its contribution is not only the assembly of city-conditioned and feature-annotated meal items, but the conversion of those items into a valid recommendation space for LLM-mediated railway service.