š®š¹šÆšµš§š· Generating multilingual instruction datasets with Magpie š¦āā¬
ā¢
18
Stefano Fiorucci PRO
anakin87
AI & ML interests
Contributing to Haystack LLM framework šļø. Language Models: orchestration, post-training, synthetic data...
Recent Activity
updated
a Space
1 day ago
anakin87/fact-checking-rocks
updated
a Space
1 day ago
anakin87/Phi-3.5-mini-ITA
updated
a Space
1 day ago
anakin87/gemma-2-2b-neogenesis-ita
Articles
Organizations
Posts
12
Post
450
Hey, it has been a while... I was busy participating in š ššš¦š¦š ššØš¦š©ššš¢šš¢šØš§!
Here's the idea: Gemma open models have a large vocabulary size (256K), so improving them for a specific language or cultural context should be pretty affordable - no need for continued pre-training.
My submission: ššš®š¹ šššØš šš§šš¬š¢š¬ - ššØš¬š-šš«šš¢š§š¢š§š ššš¦š¦š ššØš« šššš„š¢šš§ šš§š ššš²šØš§š
š Kaggle notebook: https://www.kaggle.com/code/anakin87/post-training-gemma-for-italian-and-beyond
In this notebook, I show how I improve the performance of Gemma 2 2B on Italian via Post-Training.
I believe this method is adaptable to other languages and model sizes.
šš¦šŗ ššµš¦š±š“
š Choose reference metrics
š§āš¬ Data curation for Instruction Fine Tuning: identify existing datasets + generate synthetic data
šļøāāļø Efficient Instruction Fine Tuning with Spectrum
š§āš¬ Data curation for Preference Tuning: identify existing datasets + generate synthetic data
šš Efficient Direct Preference Optimization with Spectrum
š Evaluation
š¤ Hugging Face collection (with models and datasets): anakin87/gemma-neogenesis-67824b7bf13ac9cfe091fe2e
I'm also planning a š Gemma Giveaway (on LinkedIn - https://www.linkedin.com/in/stefano-fiorucci) in the next few days - sharing techniques, datasets, and models I used for my project... so stay tuned! š»
Here's the idea: Gemma open models have a large vocabulary size (256K), so improving them for a specific language or cultural context should be pretty affordable - no need for continued pre-training.
My submission: ššš®š¹ šššØš šš§šš¬š¢š¬ - ššØš¬š-šš«šš¢š§š¢š§š ššš¦š¦š ššØš« šššš„š¢šš§ šš§š ššš²šØš§š
š Kaggle notebook: https://www.kaggle.com/code/anakin87/post-training-gemma-for-italian-and-beyond
In this notebook, I show how I improve the performance of Gemma 2 2B on Italian via Post-Training.
I believe this method is adaptable to other languages and model sizes.
šš¦šŗ ššµš¦š±š“
š Choose reference metrics
š§āš¬ Data curation for Instruction Fine Tuning: identify existing datasets + generate synthetic data
šļøāāļø Efficient Instruction Fine Tuning with Spectrum
š§āš¬ Data curation for Preference Tuning: identify existing datasets + generate synthetic data
šš Efficient Direct Preference Optimization with Spectrum
š Evaluation
š¤ Hugging Face collection (with models and datasets): anakin87/gemma-neogenesis-67824b7bf13ac9cfe091fe2e
I'm also planning a š Gemma Giveaway (on LinkedIn - https://www.linkedin.com/in/stefano-fiorucci) in the next few days - sharing techniques, datasets, and models I used for my project... so stay tuned! š»
Post
1643
Tulu 3 SFT Mixture by AllenAI is a massive, good, multilingual dataset for fine-tuning Language Models.
Unfortunately, it was missing the "language" column.
I added it using the good old fastText.
Check out the dataset here š anakin87/tulu-3-sft-mixture-with-language
Unfortunately, it was missing the "language" column.
I added it using the good old fastText.
Check out the dataset here š anakin87/tulu-3-sft-mixture-with-language
Collections
3
Datasets and models for Neogenesis: Post-training recipe for improving Gemma 2 for a specific language. Notebook: https://t.ly/iuKdy
-
ššŖš®š¹
Gemma 2 9B Neogenesis ITA
9B Italian strong model šŖ
-
2šš¤š®š¹
Gemma 2 2B Neogenesis ITA
Chat with an Italian Small Model
-
anakin87/gemma-2-2b-neogenesis-ita
Text Generation ā¢ Updated ā¢ 589 ā¢ 5 -
anakin87/gemma-2-9b-neogenesis-ita
Text Generation ā¢ Updated ā¢ 680 ā¢ 5
Starting from 2024-11-15
-
Qwen2.5 Technical Report
Paper ā¢ 2412.15115 ā¢ Published ā¢ 340 -
SPaR: Self-Play with Tree-Search Refinement to Improve Instruction-Following in Large Language Models
Paper ā¢ 2412.11605 ā¢ Published ā¢ 17 -
472š
Scaling test-time compute
-
Reverse Thinking Makes LLMs Stronger Reasoners
Paper ā¢ 2411.19865 ā¢ Published ā¢ 20
spaces
5
Running
40
šø
Fact Checking rocks!
Running
on
Zero
4
š¬š®š¹
Phi 3.5 Mini ITA
Chat with an Italian Small Model
Running
on
Zero
2
šš¤š®š¹
Gemma 2 2B Neogenesis ITA
Chat with an Italian Small Model
Running
on
Zero
ššŖš®š¹
Gemma 2 9B Neogenesis ITA
9B Italian strong model šŖ
Running
15
š»š»
Who killed Laura Palmer?
models
14
anakin87/gemma-2-2b-neogenesis-ita
Text Generation
ā¢
Updated
ā¢
589
ā¢
5
anakin87/gemma-2-9b-neogenesis-ita
Text Generation
ā¢
Updated
ā¢
680
ā¢
5
anakin87/gemma-2-2b-ita-sft
Text Generation
ā¢
Updated
ā¢
124
anakin87/Phi-3.5-mini-ITA
Text Generation
ā¢
Updated
ā¢
6.97k
ā¢
11
anakin87/yo-Llama-3-8B-Instruct
Text Generation
ā¢
Updated
ā¢
355
ā¢
7
anakin87/Llama-3-8b-ita-ties
Text Generation
ā¢
Updated
ā¢
2.85k
ā¢
3
anakin87/Llama-3-8b-ita-slerp
Text Generation
ā¢
Updated
ā¢
4.84k
ā¢
1
anakin87/Llama-3-8b-ita-ties-pro
Text Generation
ā¢
Updated
ā¢
4.79k
ā¢
1
anakin87/gemma-2b-orpo
Text Generation
ā¢
Updated
ā¢
43
ā¢
28
anakin87/gemma-2b-orpo-GGUF
Updated
ā¢
7
ā¢
7
datasets
6
anakin87/evol-dpo-ita-reranked
Viewer
ā¢
Updated
ā¢
19.8k
ā¢
17
anakin87/gemma-vs-gemma-preferences
Viewer
ā¢
Updated
ā¢
24.7k
ā¢
25
anakin87/fine-instructions-ita-70k
Viewer
ā¢
Updated
ā¢
69.9k
ā¢
15
anakin87/FineTome-single-turn-dedup
Viewer
ā¢
Updated
ā¢
83.3k
ā¢
35
anakin87/tulu-3-sft-mixture-with-language
Viewer
ā¢
Updated
ā¢
939k
ā¢
51
anakin87/medrag-pubmed-chunk
Viewer
ā¢
Updated
ā¢
15.4k
ā¢
131