TalkFa

A unified benchmark for Farsi dialogue generation and understanding

Three human-verified Farsi dialogue datasets, released with the paper TalkFa: A Unified Benchmark for Farsi Dialogue Generation and Understanding. LLMs drafted the data and native Farsi speakers reviewed and revised every dialogue. Code, guidelines and fine-tuned adapters: GitHub and the TalkFa collection.

Config	Train	Validation	Test	Task
wiki_fadial	4,084	–	100	Knowledge-grounded six-turn dialogue generation from Farsi Wikipedia
dailydialog_fa	46,946	2,471	2,601	Turn-level dialogue act (4) and emotion (7) classification
playdial_fa	1,867	99	104	Dialogue-level sentiment (negative / neutral / positive)

from datasets import load_dataset

wiki = load_dataset("Kamyar-zeinalipour/TalkFa", "wiki_fadial")
daily = load_dataset("Kamyar-zeinalipour/TalkFa", "dailydialog_fa")
plays = load_dataset("Kamyar-zeinalipour/TalkFa", "playdial_fa")

Datasets

Wiki-FaDial (4,184 dialogues, 25,104 turns). GPT-4o turned popular Farsi Wikipedia paragraphs into casual two-speaker dialogues; native speakers revised 63% of them for faithfulness and naturalness. Fields: article_title, source_list, paragraph, dialogue (list of {speaker, text}), messages (training chat).

DailyDialog-Fa (52,018 turns from 6,599 dialogues). DailyDialog (Li et al., 2017) translated into natural spoken Farsi with GPT-4o and reviewed in two stages, keeping the original labels. Fields: speaker, text, act, emotion, dialogue_index, messages_act, messages_emotion.

PlayDial-Fa (2,070 dialogues). Scenes from 30 public-domain Farsi plays (1940–2020), segmented and reformulated with GPT-4o, annotated for sentiment. Fields: title, source, characters, dialogue_type, speaker_sentiments, dialogue, sentiment, messages.

Results

Wiki-FaDial generation, base → LoRA (test set):

Model	ROUGE-L	BLEU-4	BERTCos
Llama-3.2-1B-Instruct	0.449 → 0.650	0.322 → 0.572	0.824 → 0.959
Llama-3.2-3B-Instruct	0.353 → 0.660	0.229 → 0.582	0.714 → 0.959
Mistral-7B-Instruct-v0.3	0.601 → 0.683	0.499 → 0.614	0.895 → 0.969
Llama-3.1-8B-Instruct	0.655 → 0.668	0.578 → 0.596	0.944 → 0.962
Mistral-Nemo-Instruct-2407	0.613 → 0.688	0.515 → 0.614	0.893 → 0.973
Mistral-Small-24B-Instruct-2501	0.546 → 0.687	0.434 → 0.617	0.919 → 0.972

PlayDial-Fa sentiment, macro-averaged, base → LoRA (test set):

Model	Precision	Recall	F1
Llama-3.2-1B-Instruct	0.16 → 0.33	0.33 → 0.34	0.21 → 0.23
Llama-3.2-3B-Instruct	0.33 → 0.58	0.45 → 0.52	0.38 → 0.50
Mistral-7B-Instruct-v0.3	0.56 → 0.60	0.08 → 0.58	0.13 → 0.57
Llama-3.1-8B-Instruct	0.82 → 0.64	0.18 → 0.55	0.22 → 0.55
Mistral-Nemo-Instruct-2407	0.58 → 0.63	0.54 → 0.57	0.53 → 0.57
Mistral-Small-24B-Instruct-2501	0.63 → 0.65	0.52 → 0.60	0.53 → 0.62

Fine-tuned adapters

Base model	Params	Wiki-FaDial generation	PlayDial-Fa sentiment
Llama-3.2-1B-Instruct	1B	TalkFa-WikiFaDial-Llama-3.2-1B-Instruct-LoRA	TalkFa-Sentiment-Llama-3.2-1B-Instruct-LoRA
Llama-3.2-3B-Instruct	3B	TalkFa-WikiFaDial-Llama-3.2-3B-Instruct-LoRA	TalkFa-Sentiment-Llama-3.2-3B-Instruct-LoRA
Mistral-7B-Instruct-v0.3	7B	TalkFa-WikiFaDial-Mistral-7B-Instruct-v0.3-LoRA	TalkFa-Sentiment-Mistral-7B-Instruct-v0.3-LoRA
Llama-3.1-8B-Instruct	8B	TalkFa-WikiFaDial-Llama-3.1-8B-Instruct-LoRA	TalkFa-Sentiment-Llama-3.1-8B-Instruct-LoRA
Mistral-Nemo-Instruct-2407	12B	TalkFa-WikiFaDial-Mistral-Nemo-Instruct-2407-LoRA	TalkFa-Sentiment-Mistral-Nemo-Instruct-2407-LoRA
Mistral-Small-24B-Instruct-2501	24B	TalkFa-WikiFaDial-Mistral-Small-24B-Instruct-2501-LoRA	TalkFa-Sentiment-Mistral-Small-24B-Instruct-2501-LoRA

Citation

@misc{jamshidi2026talkfa,
  title         = {TalkFa: A Unified Benchmark for Farsi Dialogue Generation and Understanding},
  author        = {Neda Jamshidi and Kamyar Zeinalipour and Fahimeh Akbari and Monica Bianchini and Marco Maggini and Marco Gori},
  year          = {2026},
  eprint        = {2609.01810},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.01810}
}

License

CC BY-NC-SA 4.0 (DailyDialog is distributed under CC BY-NC-SA 4.0; Wiki-FaDial is derived from Wikipedia).