MFMD4Instruction: Multilingual Financial Misinformation

• MFMD4Instruction is a multilingual, multitask dataset designed to standardize heterogeneous financial misinformation sources into unified instruction formats.
• It aggregates data from nine diverse sources in English, Chinese, Bengali, and Greek, enabling structured financial claim verification and fact-checking tasks.
• Serving as the training backbone for MFMDQwen, it enables effective LLM adaptation for detecting misinformation, achieving high performance on held-out benchmarks.

MFMD4Instruction is a multilingual, multitask instruction-tuning dataset introduced as the training backbone of MFMDQwen, an open-source LLM for Multilingual Financial Misinformation Detection (MFMD). It is designed to convert heterogeneous financial misinformation resources into a unified supervised fine-tuning format across English, Chinese, Greek, and Bengali, and it is paired with MFMDBench, a held-out benchmark for evaluating multilingual financial misinformation capabilities. In the paper that introduces it, MFMD4Instruction is presented as “the first instruction dataset supporting MFMD with LLMs,” with the central objective of adapting a pretrained model to financial claim verification, fact-checking, misinformation detection, and social-media manipulation detection under a shared instruction-following formulation (Liu et al., 20 Apr 2026).

1. Definition and problem setting

MFMD4Instruction is situated within the broader task of Multilingual Financial Misinformation Detection, which the paper frames as automatically determining whether a financial claim, news item, social-media post, or financially relevant statement is true, false, manipulated, entailed, refuted, supported, or unsupported, depending on the source task and label space. The motivating premise is that financial misinformation can affect market stability and investment decisions, while multilingual and culturally heterogeneous settings make the detection problem more difficult than English-only misinformation classification (Liu et al., 20 Apr 2026).

The dataset is not a generic financial corpus. It is a supervised fine-tuning resource built explicitly for instruction tuning. In the authors’ formulation, each task is represented by context-target pairs,

$D_t = \{(q_i^t, r_i^t)\}_{i=1}^{N_t},$

where $q_i^t$ contains the task description, input text, and query, and $r_i^t$ contains the target answer. The multitask objective merges the task-specific datasets and maximizes conditional likelihood:

$$\max_{\phi} \sum_{t} \sum_{i=1}^{N_t} \log \mathbf{P}_{\phi}(r_i^t \mid q_i^t).$$

This formalization is the key technical role of MFMD4Instruction: heterogeneous financial misinformation datasets are rewritten as a single instruction-following supervised fine-tuning corpus (Liu et al., 20 Apr 2026).

The paper positions the resource against prior English-centric or single-task systems and datasets, including FinFact, FinDVer, RFCBench, and FMDLlama. The claimed novelty is the conjunction of three properties: multilinguality, financial misinformation specialization, and instruction-tuning readiness for LLMs. A plausible implication is that the dataset is intended less as a benchmark in its own right than as an adaptation layer between pretrained LLMs and multilingual financial verification tasks.

2. Corpus composition and source datasets

MFMD4Instruction is built from nine existing sources spanning four languages: English, Chinese, Bengali, and Greek. The paper states that the raw corpus is compiled “based on MFMDScen and RFC,” and the summarized allocation between the supervised fine-tuning set and the held-out benchmark is given in Table 1 of the paper (Liu et al., 20 Apr 2026).

Dataset (language)	MFMD4Instruction	MFMDBench
FinDVer (EN)	554	140
CHEF (ZH)	894	238
MDFEND (ZH)	1044	265
Bengali (BN)	80	21
Global-EN (EN)	115	29
Global-ZH (ZH)	115	29
Global-BN (BN)	115	29
Global-GR (GR)	115	29
RFC (EN)	1805	1652
Total	4837	2432

The total pool is 7,269 examples, of which 4,837 belong to MFMD4Instruction and 2,432 to MFMDBench. The source tasks are heterogeneous. FinDVer is financial claim verification with entailed/refuted labels; CHEF is evidence-based fact-checking with supported/refuted/not enough information; MDFEND is real/fake content detection on Weibo; BanMANI is manipulated versus not manipulated social-media detection; Global4Languages provides aligned true/false claim verification across English, Chinese, Bengali, and Greek; and RFC-Bench contributes reference-free counterfactual financial misinformation examples (Liu et al., 20 Apr 2026).

The paper explicitly notes several inconsistencies between prose descriptions and the summary table. FinDVer is described in the narrative as having 700 instances, but the table reports 694. CHEF is described as having 1,188 samples, while the table reports 1,132. MDFEND is described as having 1,321 posts, while the table reports 1,309. RFC-Bench is described as containing 1,826 original-perturbed paragraph pairs, whereas the table reports a total of 3,457 examples. These discrepancies are not explained in the paper. A cautious reading is therefore that some filtering or preprocessing occurred, but the exact criteria are not documented.

3. Instruction schema and task heterogeneity

MFMD4Instruction standardizes source datasets by converting them into structured prompts containing a task description and the relevant input fields. The paper reproduces the prompt templates in its appendix. For Global4Languages, the template is “Determine whether the claim is ‘True’ or ‘False’.” For FinDVer, the model is asked to assess whether a claim is entailed or refuted based on a relevant financial report. For MDFEND, the task is to determine whether content is “real” or “false.” For CHEF, the output space is Supported, Refuted, or Not enough information based on provided evidence. For BanMANI, the model must output MANI or NO_MANI based on original news and a social-media post. For RFC, the prompt asks the model to determine whether financial information is true or false (Liu et al., 20 Apr 2026).

This design choice is important because MFMD4Instruction does not collapse all tasks into a single unified label ontology. Instead, it preserves the output conventions of the source datasets and unifies them at the level of instruction-following generation. The paper states that each sample is transformed into a structured record containing “(1) a task description specifying the detection task and expected output format, (2) the input fields consisting of the claim or content together with any supporting evidence or contextual information.” The resulting corpus is therefore multitask in a strong sense: binary claim verification, entailment-style verification, three-way evidence-based fact-checking, social-media manipulation detection, and real/fake content detection coexist within one supervised fine-tuning format (Liu et al., 20 Apr 2026).

The labels are inherited from the source datasets rather than newly reannotated. This has two consequences. First, annotation methodology is source-dependent. Second, the corpus preserves cross-dataset heterogeneity in genre, evidence structure, and label semantics. The paper reports no unified reannotation pass, no combined inter-annotator agreement, and no additional human verification after conversion to instruction format. RFC-Bench is the one source for which the paper explicitly mentions multi-stage expert review and dual-annotator evaluation, but that quality-control statement is confined to RFC rather than the entire assembled corpus.

4. Role in MFMDQwen training

Within the MFMDQwen system, MFMD4Instruction is the supervised fine-tuning dataset used to adapt a general-purpose base model to multilingual financial misinformation detection. The paper states that the model is built on Qwen-3-8B and that training proceeds in two stages, “SFT followed by RL.” The only explicit optimization formula given in the paper is the supervised fine-tuning objective over the merged instruction-conditioned task datasets (Liu et al., 20 Apr 2026).

The reported hyperparameters for the MFMD4Instruction-based stage are specific. The learning rate is $1\times10^{-5}$, the warmup ratio is 0.1, training runs for 2 epochs, and the batch size is 128. The implementation uses DeepSpeed ZeRO-3 with CPU parameter offloading, a maximum input sequence length of 24k tokens, a maximum output length of 8k tokens, and full-parameter training on four NVIDIA L40s GPUs with 48 GB each. The paper does not provide corresponding details for the RL stage, nor does it isolate the contribution of RL relative to SFT (Liu et al., 20 Apr 2026).

The model nomenclature in the paper is slightly unstable. One passage states, “We then built MFMD-R based on Qwen-3-8B using the MFMD4Instruction dataset,” whereas the title and results table emphasize MFMDQwen. The text provided does not clarify whether MFMD-R is an internal name for the same model or a distinct variant. What is clear is that MFMD4Instruction is the dataset used for the SFT stage that underlies the reported downstream results.

5. Empirical evidence through MFMDBench

The effectiveness of MFMD4Instruction is evaluated indirectly through MFMDBench, the held-out benchmark constructed from the same source ecosystem. MFMDBench covers 9 tasks across 4 languages, and the main reported metrics are Accuracy and Macro-F1, with the paper stating that the analysis focuses on F1. On this benchmark, MFMDQwen achieves the best average F1 among the listed systems, at 0.818 (Liu et al., 20 Apr 2026).

The next-best averages reported in the table are lower: Qwen3-32b-R reaches 0.674, Llama3.3-70b reaches 0.663, Qwen2.5-72b reaches 0.662, and FMDLlama reaches 0.293. The paper further states that MFMDQwen achieves top performance on 6 of the 9 benchmarks. Its task-wise F1 scores are reported as 0.758 on GlobalGr, 0.803 on GlobalBe, 0.710 on GlobalCh, 0.756 on GlobalEn, 0.843 on CHEF, 0.913 on MDFEND, 0.952 on Bengali, 0.673 on FinDVer, and 0.954 on RFC. The strongest gains are reported on Chinese datasets, especially CHEF and MDFEND, and the paper attributes this to the multilingual supervised fine-tuning strategy enabled by MFMD4Instruction (Liu et al., 20 Apr 2026).

At the same time, the evidence is not an ablation in the strict sense. The paper does not compare the same base model with and without MFMD4Instruction, and it does not separately evaluate multilingual versus monolingual training variants. The empirical claim is therefore comparative rather than isolating: the MFMDQwen system trained on MFMD4Instruction outperforms general LLM baselines and the prior domain-specific FMDLlama on the held-out multilingual benchmark.

The paper also includes an error analysis that identifies five recurring failure types: label polarity reversal and prior-driven guessing, numerical reasoning and arithmetic errors, table parsing and financial statement grounding errors, partial evidence matching that misses a critical incorrect detail, and manipulation detection failures under subtle lexical mismatch. This suggests that MFMD4Instruction improves multilingual domain adaptation but does not eliminate deeper reasoning bottlenecks in financial verification.

6. Limitations, interpretation, and relation to adjacent benchmark work

The paper emphasizes several limitations of MFMD4Instruction. The dataset is imbalanced across languages, with English and Chinese contributing most of the data while Bengali and especially Greek remain small. The labels are inherited from heterogeneous source datasets, so annotation policy and genre vary by subtask. The paper provides no separate train/validation/test split within MFMD4Instruction, no validation set for instruction tuning, and no documented unified preprocessing policy to explain the source-count discrepancies. It also does not report a unified reannotation pass or inter-annotator agreement for the assembled instruction corpus (Liu et al., 20 Apr 2026).

These characteristics have direct methodological implications. The resource is best understood as an instruction-format unification layer rather than a newly annotated, ontology-standardized multilingual benchmark. This suggests strengths in practical supervised adaptation, but also a dependence on the quality, balance, and latent biases of the original sources. It also suggests that evaluation on MFMDBench, while useful, remains within the same source ecosystem from which MFMD4Instruction is assembled.

In the broader instruction-following literature, MFMD4Instruction occupies a different position from multilingual evaluation frameworks such as M-IFEval, which extends objective instruction-following assessment to French, Japanese, and Spanish, and from domain-general code benchmarks such as MultiCodeIF, which targets fine-grained code instruction following with hierarchical constraints and feedback-driven refinement (Dussolle et al., 7 Feb 2025, Duan et al., 1 Jul 2025). MFMD4Instruction is narrower in domain but more directly operational: it is a supervised fine-tuning dataset specialized for multilingual financial misinformation detection rather than a general instruction-following benchmark.

Its significance lies in that specialization. It combines multilingual coverage, task heterogeneity, and instruction-tuning format in a domain where small factual distortions can have outsized consequences. The paper therefore presents MFMD4Instruction not as a universal instruction-following resource, but as a domain-specific infrastructure for adapting LLMs to multilingual financial verification, claim checking, and manipulation detection under a shared generative training objective (Liu et al., 20 Apr 2026).