{ "id": 663357, "title": "Other Public Data", "url": "/competitions/deep-past-initiative-machine-translation/discussion/663357", "author_name": "", "post_date": "2025-12-17T15:27:36.488702300Z", "votes": 31, "comment_count": 9, "views": 0, "content": "

This might be useful:

\n

https://www.kaggle.com/datasets/manwithacat/oracc-akkadian-english-parallel-corpus

\n

I’m sharing ORACC Akkadian–English Parallel Corpus with the community as a clean, ready-to-use set of aligned Akkadian transliteration → English pairs extracted from ORACC. The material here is drawn from major ORACC projects focused on royal inscriptions and state/administrative archives (RIAo, RINAP, RIBo, SAAo), so it is not a purpose-built corpus of Old Assyrian / Old Akkadian merchant records. In other words: this dataset is best treated as a broad, high-quality Akkadian–English parallel resource—particularly relevant to Neo-Assyrian / Neo-Babylonian-style language and genres—useful for pretraining, baseline NMT, alignment experiments, and tooling, rather than a direct proxy for the Kaneš trade letters.

\n

ORACC Akkadian–English Parallel Corpus

\n

This dataset contains parallel pairs of Akkadian transliteration (source) and English translation (target), extracted and aligned from the Open Richly Annotated Cuneiform Corpus (ORACC).

\n

Source Projects

\n

RIAo (Royal Inscriptions of Assyria online): Assyrian royal inscriptions ~2500-609 BCE\nRINAP (Royal Inscriptions of the Neo-Assyrian Period): Neo-Assyrian royal texts 744-609 BCE\nRIBo (Royal Inscriptions of Babylonia online): Babylonian royal inscriptions ~2500-539 BCE\nSAAo (State Archives of Assyria online): Letters and administrative texts 911-612 BCE

", "messages": [ { "id": "3378131", "postDate": "12/17/2025 15:27:36", "content": "

This might be useful:

\n

https://www.kaggle.com/datasets/manwithacat/oracc-akkadian-english-parallel-corpus

\n

I’m sharing ORACC Akkadian–English Parallel Corpus with the community as a clean, ready-to-use set of aligned Akkadian transliteration → English pairs extracted from ORACC. The material here is drawn from major ORACC projects focused on royal inscriptions and state/administrative archives (RIAo, RINAP, RIBo, SAAo), so it is not a purpose-built corpus of Old Assyrian / Old Akkadian merchant records. In other words: this dataset is best treated as a broad, high-quality Akkadian–English parallel resource—particularly relevant to Neo-Assyrian / Neo-Babylonian-style language and genres—useful for pretraining, baseline NMT, alignment experiments, and tooling, rather than a direct proxy for the Kaneš trade letters.

\n

ORACC Akkadian–English Parallel Corpus

\n

This dataset contains parallel pairs of Akkadian transliteration (source) and English translation (target), extracted and aligned from the Open Richly Annotated Cuneiform Corpus (ORACC).

\n

Source Projects

\n

RIAo (Royal Inscriptions of Assyria online): Assyrian royal inscriptions ~2500-609 BCE\nRINAP (Royal Inscriptions of the Neo-Assyrian Period): Neo-Assyrian royal texts 744-609 BCE\nRIBo (Royal Inscriptions of Babylonia online): Babylonian royal inscriptions ~2500-539 BCE\nSAAo (State Archives of Assyria online): Letters and administrative texts 911-612 BCE

", "rawMarkdown": "This might be useful:\n\nhttps://www.kaggle.com/datasets/manwithacat/oracc-akkadian-english-parallel-corpus\n\nI’m sharing ORACC Akkadian–English Parallel Corpus with the community as a clean, ready-to-use set of aligned Akkadian transliteration → English pairs extracted from ORACC. The material here is drawn from major ORACC projects focused on royal inscriptions and state/administrative archives (RIAo, RINAP, RIBo, SAAo), so it is not a purpose-built corpus of Old Assyrian / Old Akkadian merchant records. In other words: this dataset is best treated as a broad, high-quality Akkadian–English parallel resource—particularly relevant to Neo-Assyrian / Neo-Babylonian-style language and genres—useful for pretraining, baseline NMT, alignment experiments, and tooling, rather than a direct proxy for the Kaneš trade letters.\n\nORACC Akkadian–English Parallel Corpus\n\nThis dataset contains parallel pairs of Akkadian transliteration (source) and English translation (target), extracted and aligned from the Open Richly Annotated Cuneiform Corpus (ORACC).\n\nSource Projects\n\nRIAo (Royal Inscriptions of Assyria online): Assyrian royal inscriptions ~2500-609 BCE\nRINAP (Royal Inscriptions of the Neo-Assyrian Period): Neo-Assyrian royal texts 744-609 BCE\nRIBo (Royal Inscriptions of Babylonia online): Babylonian royal inscriptions ~2500-539 BCE\nSAAo (State Archives of Assyria online): Letters and administrative texts 911-612 BCE", "votes": null }, { "id": "3378237", "postDate": "12/17/2025 19:34:11", "content": "

How very resourceful you are. As a matter of fact, I am quite familiar with this dataset, and I would urge caution in using it, mainly because these texts from Oracc are very different from the Old Assyrian archives, in terms of time period and genre. First, there are no other Old Assyrian texts in the Oracc data, and only very few texts of a contemporary time period. The vast majority of these texts are from the first millennium, and therefore are about 1000 years later in time, which makes a huge impact on the use of the language. Second, we found in using this dataset that there were a large number of lexical lists, which were used in antiquity to record Akkadian to Sumerian equivalencies. Many of these lexical lists are also quite late (i.e. Late Babylonian), and therefore they will contain words which have no bearing on the Old Assyrian texts or the ways in which they wrote / spoke.

\n

I understand that it is desirable to find as much data as possible for this challenge, but consider this. If you were building a Machine Translation model for modern English, would you also include Middle English texts, or maybe Old English lexicons? Probably not, but that's just my point of view.

", "rawMarkdown": "How very resourceful you are. As a matter of fact, I am quite familiar with this dataset, and I would urge caution in using it, mainly because these texts from Oracc are very different from the Old Assyrian archives, in terms of time period and genre. First, there are no other Old Assyrian texts in the Oracc data, and only very few texts of a contemporary time period. The vast majority of these texts are from the first millennium, and therefore are about 1000 years later in time, which makes a huge impact on the use of the language. Second, we found in using this dataset that there were a large number of lexical lists, which were used in antiquity to record Akkadian to Sumerian equivalencies. Many of these lexical lists are also quite late (i.e. Late Babylonian), and therefore they will contain words which have no bearing on the Old Assyrian texts or the ways in which they wrote / spoke. \n\nI understand that it is desirable to find as much data as possible for this challenge, but consider this. If you were building a Machine Translation model for modern English, would you also include Middle English texts, or maybe Old English lexicons? Probably not, but that's just my point of view.", "votes": null }, { "id": "3378252", "postDate": "12/17/2025 20:15:42", "content": "

I felt it was the least worst option available to try and add something to my model that was vaguely Akkadian. I suppose using your analogy if I was trying to build a machine translation model for modern English and all I had was middle English then maybe I could at least extract some nouns or names or guess at some grammar rules from it.

\n

Sumer Is Icumen In…

", "rawMarkdown": "I felt it was the least worst option available to try and add something to my model that was vaguely Akkadian. I suppose using your analogy if I was trying to build a machine translation model for modern English and all I had was middle English then maybe I could at least extract some nouns or names or guess at some grammar rules from it.\n\nSumer Is Icumen In…", "votes": null }, { "id": "3378267", "postDate": "12/17/2025 20:27:46", "content": "

I understand, in fact I tried this myself. I found that the genre is lost / muddled in translation by including all of this royal inscription text. Some words are generic enough that they are used in many different contexts, so when there's a bunch of examples from the first millennium of words with contexts like kings, conquests, priests, sacrifices, and the like, it creates a very different vector for words which in Old Assyrian are used with economic contexts (almost exclusively). To put it in another way, the sentence in Old Assyrian: \"he should send the money to me, and I will pay the tax\", under the influence of Neo-Assyrian or Neo-Babylonian royal inscriptions might come out as \"he shall send me the tithes, and I will make an offering\". This is a made up example, but it is very apparent in the AICC translations from 2023 of the CDLI (https://aicuneiform.com/).

", "rawMarkdown": "I understand, in fact I tried this myself. I found that the genre is lost / muddled in translation by including all of this royal inscription text. Some words are generic enough that they are used in many different contexts, so when there's a bunch of examples from the first millennium of words with contexts like kings, conquests, priests, sacrifices, and the like, it creates a very different vector for words which in Old Assyrian are used with economic contexts (almost exclusively). To put it in another way, the sentence in Old Assyrian: \"he should send the money to me, and I will pay the tax\", under the influence of Neo-Assyrian or Neo-Babylonian royal inscriptions might come out as \"he shall send me the tithes, and I will make an offering\". This is a made up example, but it is very apparent in the AICC translations from 2023 of the CDLI (https://aicuneiform.com/).", "votes": null }, { "id": "3378275", "postDate": "12/17/2025 20:45:02", "content": "

Out of interest , I’ve got a plan and I’m wondering if I’m covering territory that’s already been tried before :

\n

I’m proposing to try, very explicitly as an experimental modelling convenience rather than a claim about historical phonology, to introduce a parallel, highly normalised representation of the Akkadian input in which syllabic spellings are partially collapsed toward an abstract consonantal layer—either by stripping vowels entirely or by mapping consonants onto coarse articulatory classes (e.g. labial, dental/alveolar, sibilant, guttural)—and to feed this alongside the original transliteration rather than in place of it.

\n

I’m not sure if that’s a silly idea or not, but I’m thinking cuneiform writing system might (does?) encode phonemic material in a deliberately lossy, syllabic way which creates extreme “surface variability” over underlying lexical and morphological regularities a model would ideally learn.

\n

Anyway, worst case scenario, I burn a few GPU hours testing it out.

", "rawMarkdown": "Out of interest , I’ve got a plan and I’m wondering if I’m covering territory that’s already been tried before :\n\n I’m proposing to try, very explicitly as an experimental modelling convenience rather than a claim about historical phonology, to introduce a parallel, highly normalised representation of the Akkadian input in which syllabic spellings are partially collapsed toward an abstract consonantal layer—either by stripping vowels entirely or by mapping consonants onto coarse articulatory classes (e.g. labial, dental/alveolar, sibilant, guttural)—and to feed this alongside the original transliteration rather than in place of it. \n\nI’m not sure if that’s a silly idea or not, but I’m thinking cuneiform writing system might (does?) encode phonemic material in a deliberately lossy, syllabic way which creates extreme “surface variability” over underlying lexical and morphological regularities a model would ideally learn.\n\nAnyway, worst case scenario, I burn a few GPU hours testing it out.", "votes": null }, { "id": "3378342", "postDate": "12/17/2025 23:20:18", "content": "

That is not a silly idea at all, in fact it makes good sense when you consider how tokenization works. The hyphens certainly pose an issue for subword tokenization, so normalizing the transliteration could be a good work-around for getting a better tokenization, and perhaps better embeddings too, depending on the size of the model.

", "rawMarkdown": "That is not a silly idea at all, in fact it makes good sense when you consider how tokenization works. The hyphens certainly pose an issue for subword tokenization, so normalizing the transliteration could be a good work-around for getting a better tokenization, and perhaps better embeddings too, depending on the size of the model.", "votes": null }, { "id": "3384424", "postDate": "01/01/2026 08:33:03", "content": "
\n

I’m proposing to try, very explicitly as an experimental modelling convenience rather than a claim about historical phonology, to introduce a parallel, highly normalised representation of the Akkadian input in which syllabic spellings are partially collapsed toward an abstract consonantal layer—either by stripping vowels entirely or by mapping consonants onto coarse articulatory classes (e.g. labial, dental/alveolar, sibilant, guttural)—and to feed this alongside the original transliteration rather than in place of it.

\n
\n

@manwithacat, is this normalized transliteration idea related to your thought of using ORACC royal inscriptions and state/administrative archives as data?

", "rawMarkdown": "> I’m proposing to try, very explicitly as an experimental modelling convenience rather than a claim about historical phonology, to introduce a parallel, highly normalised representation of the Akkadian input in which syllabic spellings are partially collapsed toward an abstract consonantal layer—either by stripping vowels entirely or by mapping consonants onto coarse articulatory classes (e.g. labial, dental/alveolar, sibilant, guttural)—and to feed this alongside the original transliteration rather than in place of it. \n\n@manwithacat, is this normalized transliteration idea related to your thought of using ORACC royal inscriptions and state/administrative archives as data?", "votes": null }, { "id": "3386486", "postDate": "01/05/2026 11:13:51", "content": "

It was, but as the competition host noted above there’s a lot that is lost in translation. Also by their nature these types of inscriptions/proclamations are repetitive which is kind of counter-productive for model training.

\n

My best score is currently using the public model trained by @jeanjean111

", "rawMarkdown": "It was, but as the competition host noted above there’s a lot that is lost in translation. Also by their nature these types of inscriptions/proclamations are repetitive which is kind of counter-productive for model training.\n\nMy best score is currently using the public model trained by @jeanjean111", "votes": null }, { "id": "3416431", "postDate": "03/02/2026 21:14:57", "content": "

I should mention, there is a very large corpus of Oracc texts which I provided in a zenodo repository here as well (https://zenodo.org/records/10794626). That said, the same warnings apply here. These include many different genres and languages, and almost none of them are Old Assyrian texts. Use at your own risk.

", "rawMarkdown": "I should mention, there is a very large corpus of Oracc texts which I provided in a zenodo repository here as well (https://zenodo.org/records/10794626). That said, the same warnings apply here. These include many different genres and languages, and almost none of them are Old Assyrian texts. Use at your own risk.", "votes": null }, { "id": "3416433", "postDate": "03/02/2026 21:22:01", "content": "

This website that is much more aligned in terms of time period (Old Babylonian): https://archibab.fr/textes?trloo=false&dmin=2020&dmax=0&o=ref,id&page=1&size=25&count=32625

", "rawMarkdown": "This website that is much more aligned in terms of time period (Old Babylonian): https://archibab.fr/textes?trloo=false&dmin=2020&dmax=0&o=ref,id&page=1&size=25&count=32625", "votes": null } ], "comments": [ { "id": 3378237, "author_name": "deeppast", "author_url": "", "post_date": "12/17/2025 19:34:11", "content": "

How very resourceful you are. As a matter of fact, I am quite familiar with this dataset, and I would urge caution in using it, mainly because these texts from Oracc are very different from the Old Assyrian archives, in terms of time period and genre. First, there are no other Old Assyrian texts in the Oracc data, and only very few texts of a contemporary time period. The vast majority of these texts are from the first millennium, and therefore are about 1000 years later in time, which makes a huge impact on the use of the language. Second, we found in using this dataset that there were a large number of lexical lists, which were used in antiquity to record Akkadian to Sumerian equivalencies. Many of these lexical lists are also quite late (i.e. Late Babylonian), and therefore they will contain words which have no bearing on the Old Assyrian texts or the ways in which they wrote / spoke.

\n

I understand that it is desirable to find as much data as possible for this challenge, but consider this. If you were building a Machine Translation model for modern English, would you also include Middle English texts, or maybe Old English lexicons? Probably not, but that's just my point of view.

", "votes": null, "replies": [ { "id": 3378252, "author_name": "manwithacat", "author_url": "", "post_date": "12/17/2025 20:15:42", "content": "

I felt it was the least worst option available to try and add something to my model that was vaguely Akkadian. I suppose using your analogy if I was trying to build a machine translation model for modern English and all I had was middle English then maybe I could at least extract some nouns or names or guess at some grammar rules from it.

\n

Sumer Is Icumen In…

", "votes": null, "replies": [ { "id": 3378267, "author_name": "deeppast", "author_url": "", "post_date": "12/17/2025 20:27:46", "content": "

I understand, in fact I tried this myself. I found that the genre is lost / muddled in translation by including all of this royal inscription text. Some words are generic enough that they are used in many different contexts, so when there's a bunch of examples from the first millennium of words with contexts like kings, conquests, priests, sacrifices, and the like, it creates a very different vector for words which in Old Assyrian are used with economic contexts (almost exclusively). To put it in another way, the sentence in Old Assyrian: \"he should send the money to me, and I will pay the tax\", under the influence of Neo-Assyrian or Neo-Babylonian royal inscriptions might come out as \"he shall send me the tithes, and I will make an offering\". This is a made up example, but it is very apparent in the AICC translations from 2023 of the CDLI (https://aicuneiform.com/).

", "votes": null, "replies": [ { "id": 3378275, "author_name": "manwithacat", "author_url": "", "post_date": "12/17/2025 20:45:02", "content": "

Out of interest , I’ve got a plan and I’m wondering if I’m covering territory that’s already been tried before :

\n

I’m proposing to try, very explicitly as an experimental modelling convenience rather than a claim about historical phonology, to introduce a parallel, highly normalised representation of the Akkadian input in which syllabic spellings are partially collapsed toward an abstract consonantal layer—either by stripping vowels entirely or by mapping consonants onto coarse articulatory classes (e.g. labial, dental/alveolar, sibilant, guttural)—and to feed this alongside the original transliteration rather than in place of it.

\n

I’m not sure if that’s a silly idea or not, but I’m thinking cuneiform writing system might (does?) encode phonemic material in a deliberately lossy, syllabic way which creates extreme “surface variability” over underlying lexical and morphological regularities a model would ideally learn.

\n

Anyway, worst case scenario, I burn a few GPU hours testing it out.

", "votes": null, "replies": [ { "id": 3378342, "author_name": "deeppast", "author_url": "", "post_date": "12/17/2025 23:20:18", "content": "

That is not a silly idea at all, in fact it makes good sense when you consider how tokenization works. The hyphens certainly pose an issue for subword tokenization, so normalizing the transliteration could be a good work-around for getting a better tokenization, and perhaps better embeddings too, depending on the size of the model.

", "votes": null, "replies": [] }, { "id": 3384424, "author_name": "jofatmofn", "author_url": "", "post_date": "01/01/2026 08:33:03", "content": "
\n

I’m proposing to try, very explicitly as an experimental modelling convenience rather than a claim about historical phonology, to introduce a parallel, highly normalised representation of the Akkadian input in which syllabic spellings are partially collapsed toward an abstract consonantal layer—either by stripping vowels entirely or by mapping consonants onto coarse articulatory classes (e.g. labial, dental/alveolar, sibilant, guttural)—and to feed this alongside the original transliteration rather than in place of it.

\n
\n

@manwithacat, is this normalized transliteration idea related to your thought of using ORACC royal inscriptions and state/administrative archives as data?

", "votes": null, "replies": [ { "id": 3386486, "author_name": "manwithacat", "author_url": "", "post_date": "01/05/2026 11:13:51", "content": "

It was, but as the competition host noted above there’s a lot that is lost in translation. Also by their nature these types of inscriptions/proclamations are repetitive which is kind of counter-productive for model training.

\n

My best score is currently using the public model trained by @jeanjean111

", "votes": null, "replies": [] } ] } ] } ] } ] } ] }, { "id": 3416431, "author_name": "deeppast", "author_url": "", "post_date": "03/02/2026 21:14:57", "content": "

I should mention, there is a very large corpus of Oracc texts which I provided in a zenodo repository here as well (https://zenodo.org/records/10794626). That said, the same warnings apply here. These include many different genres and languages, and almost none of them are Old Assyrian texts. Use at your own risk.

", "votes": null, "replies": [ { "id": 3416433, "author_name": "deeppast", "author_url": "", "post_date": "03/02/2026 21:22:01", "content": "

This website that is much more aligned in terms of time period (Old Babylonian): https://archibab.fr/textes?trloo=false&dmin=2020&dmax=0&o=ref,id&page=1&size=25&count=32625

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3378131": "This might be useful:\n\nhttps://www.kaggle.com/datasets/manwithacat/oracc-akkadian-english-parallel-corpus\n\nI’m sharing ORACC Akkadian–English Parallel Corpus with the community as a clean, ready-to-use set of aligned Akkadian transliteration → English pairs extracted from ORACC. The material here is drawn from major ORACC projects focused on royal inscriptions and state/administrative archives (RIAo, RINAP, RIBo, SAAo), so it is not a purpose-built corpus of Old Assyrian / Old Akkadian merchant records. In other words: this dataset is best treated as a broad, high-quality Akkadian–English parallel resource—particularly relevant to Neo-Assyrian / Neo-Babylonian-style language and genres—useful for pretraining, baseline NMT, alignment experiments, and tooling, rather than a direct proxy for the Kaneš trade letters.\n\nORACC Akkadian–English Parallel Corpus\n\nThis dataset contains parallel pairs of Akkadian transliteration (source) and English translation (target), extracted and aligned from the Open Richly Annotated Cuneiform Corpus (ORACC).\n\nSource Projects\n\nRIAo (Royal Inscriptions of Assyria online): Assyrian royal inscriptions ~2500-609 BCE\nRINAP (Royal Inscriptions of the Neo-Assyrian Period): Neo-Assyrian royal texts 744-609 BCE\nRIBo (Royal Inscriptions of Babylonia online): Babylonian royal inscriptions ~2500-539 BCE\nSAAo (State Archives of Assyria online): Letters and administrative texts 911-612 BCE", "3378237": "How very resourceful you are. As a matter of fact, I am quite familiar with this dataset, and I would urge caution in using it, mainly because these texts from Oracc are very different from the Old Assyrian archives, in terms of time period and genre. First, there are no other Old Assyrian texts in the Oracc data, and only very few texts of a contemporary time period. The vast majority of these texts are from the first millennium, and therefore are about 1000 years later in time, which makes a huge impact on the use of the language. Second, we found in using this dataset that there were a large number of lexical lists, which were used in antiquity to record Akkadian to Sumerian equivalencies. Many of these lexical lists are also quite late (i.e. Late Babylonian), and therefore they will contain words which have no bearing on the Old Assyrian texts or the ways in which they wrote / spoke. \n\nI understand that it is desirable to find as much data as possible for this challenge, but consider this. If you were building a Machine Translation model for modern English, would you also include Middle English texts, or maybe Old English lexicons? Probably not, but that's just my point of view.", "3378252": "I felt it was the least worst option available to try and add something to my model that was vaguely Akkadian. I suppose using your analogy if I was trying to build a machine translation model for modern English and all I had was middle English then maybe I could at least extract some nouns or names or guess at some grammar rules from it.\n\nSumer Is Icumen In…", "3378267": "I understand, in fact I tried this myself. I found that the genre is lost / muddled in translation by including all of this royal inscription text. Some words are generic enough that they are used in many different contexts, so when there's a bunch of examples from the first millennium of words with contexts like kings, conquests, priests, sacrifices, and the like, it creates a very different vector for words which in Old Assyrian are used with economic contexts (almost exclusively). To put it in another way, the sentence in Old Assyrian: \"he should send the money to me, and I will pay the tax\", under the influence of Neo-Assyrian or Neo-Babylonian royal inscriptions might come out as \"he shall send me the tithes, and I will make an offering\". This is a made up example, but it is very apparent in the AICC translations from 2023 of the CDLI (https://aicuneiform.com/).", "3378275": "Out of interest , I’ve got a plan and I’m wondering if I’m covering territory that’s already been tried before :\n\n I’m proposing to try, very explicitly as an experimental modelling convenience rather than a claim about historical phonology, to introduce a parallel, highly normalised representation of the Akkadian input in which syllabic spellings are partially collapsed toward an abstract consonantal layer—either by stripping vowels entirely or by mapping consonants onto coarse articulatory classes (e.g. labial, dental/alveolar, sibilant, guttural)—and to feed this alongside the original transliteration rather than in place of it. \n\nI’m not sure if that’s a silly idea or not, but I’m thinking cuneiform writing system might (does?) encode phonemic material in a deliberately lossy, syllabic way which creates extreme “surface variability” over underlying lexical and morphological regularities a model would ideally learn.\n\nAnyway, worst case scenario, I burn a few GPU hours testing it out.", "3378342": "That is not a silly idea at all, in fact it makes good sense when you consider how tokenization works. The hyphens certainly pose an issue for subword tokenization, so normalizing the transliteration could be a good work-around for getting a better tokenization, and perhaps better embeddings too, depending on the size of the model.", "3384424": "> I’m proposing to try, very explicitly as an experimental modelling convenience rather than a claim about historical phonology, to introduce a parallel, highly normalised representation of the Akkadian input in which syllabic spellings are partially collapsed toward an abstract consonantal layer—either by stripping vowels entirely or by mapping consonants onto coarse articulatory classes (e.g. labial, dental/alveolar, sibilant, guttural)—and to feed this alongside the original transliteration rather than in place of it. \n\n@manwithacat, is this normalized transliteration idea related to your thought of using ORACC royal inscriptions and state/administrative archives as data?", "3386486": "It was, but as the competition host noted above there’s a lot that is lost in translation. Also by their nature these types of inscriptions/proclamations are repetitive which is kind of counter-productive for model training.\n\nMy best score is currently using the public model trained by @jeanjean111", "3416431": "I should mention, there is a very large corpus of Oracc texts which I provided in a zenodo repository here as well (https://zenodo.org/records/10794626). That said, the same warnings apply here. These include many different genres and languages, and almost none of them are Old Assyrian texts. Use at your own risk.", "3416433": "This website that is much more aligned in terms of time period (Old Babylonian): https://archibab.fr/textes?trloo=false&dmin=2020&dmax=0&o=ref,id&page=1&size=25&count=32625" }, "source": "meta" }