AI & ML interests
computational linguistics, natural language processing
Recent Activity
Papers
Forecasting Downstream Performance of LLMs With Proxy Metrics
Structured Distillation of Web Agent Capabilities Enables Generalization
Best open African LLM
-
AfriqueLLM: How Data Mixing and Model Architecture Impact Continued Pre-training for African Languages
Paper • 2601.06395 • Published • 5 -
McGill-NLP/AfriqueQwen-14B
Text Generation • 15B • Updated • 2.32k • • 4 -
McGill-NLP/AfriqueQwen-8B
Text Generation • 8B • Updated • 1.58k • • 2 -
McGill-NLP/AfriqueQwen3.5-4B-50Langs
Text Generation • 5B • Updated • 399 • 6
Generative Embeddings from Large Language Models
INJONGO: A Multicultural Intent Detection and Slot-filling Dataset for 16 African Languages
-
McGill-NLP/AfroXLMR-large-76L-Injongo-intent
Text Classification • 0.6B • Updated • 3 -
McGill-NLP/AfroXLMR-large-76L-Injongo-slot
Token Classification • 0.6B • Updated • 4 -
McGill-NLP/gemma-2-9b-it-Injongo-intent
Text Generation • 9B • Updated • 3 -
McGill-NLP/gemma-2-9b-it-Injongo-slot
Text Generation • 9B • Updated • 2
Datasets used for the OLMo experiments in the "Not All Data are Unlearned Equally" paper https://arxiv.org/abs/2504.05058
Generate challenging synthetic data to evaluate LLMs
https://mcgill-nlp.github.io/weblinx
https://mcgill-nlp.github.io/weblinx
Models and data from "Structured Distillation of Web Agent Capabilities Enables Generalization" (arXiv:2604.07776)
-
Structured Distillation of Web Agent Capabilities Enables Generalization
Paper • 2604.07776 • Published • 23 -
McGill-NLP/A3-Qwen3.5-9B
Image-Text-to-Text • 9B • Updated • 283 • 6 -
McGill-NLP/A3-Qwen3.5-4B
Image-Text-to-Text • 5B • Updated • 91 • 2 -
McGill-NLP/A3-Qwen3.5-2B
Image-Text-to-Text • 3B • Updated • 31 • 2
Pre-computed contextual text embeddings for interpreting LLM/VLM hidden states. Use with: pip install latentlens
Reformulating the RL of reasoning LLMs through Markovian Thinking paradigm.
-
AgentRewardBench: Evaluating Automatic Evaluations of Web Agent Trajectories
Paper • 2504.08942 • Published • 29 -
McGill-NLP/agent-reward-bench
Viewer • Updated • 1.41k • 7.35k • 4 -
Agent Reward Bench Demo
💻5Explore agent trajectories and judgments in web benchmarks
-
Agent Reward Bench Leaderboard
🥇3Leaderboard for AgentRewardBench
-
McGill-NLP/LLM2Vec-Meta-Llama-32-3B-Instruct-mntp-supervised
Updated -
McGill-NLP/LLM2Vec-Meta-Llama-31-8B-Instruct-mntp-supervised
Sentence Similarity • Updated • 317 • 5 -
McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised
Sentence Similarity • Updated • 112k • 52 -
McGill-NLP/LLM2Vec-Mistral-7B-Instruct-v2-mntp-supervised
Sentence Similarity • Updated • 270 • 13
Repository: https://github.com/McGill-NLP/AURORA
mcgill-nlp.github.io/statcan-dialogue-dataset
-
The StatCan Dialogue Dataset: Retrieving Data Tables through Conversations with Genuine Intents
Paper • 2304.01412 • Published • 2 -
McGill-NLP/statcan-dialogue-dataset
Preview • Updated • 4 • 7 -
McGill-NLP/dpr-statcan-conversation_encoder-title
Feature Extraction • 0.1B • Updated • 7 -
McGill-NLP/tapas-statcan-large-conversation_encoder-cell_tokens
Feature Extraction • Updated • 4
-
Back-Training excels Self-Training at Unsupervised Domain Adaptation of Question Generation and Passage Retrieval
Paper • 2104.08801 • Published • 1 -
McGill-NLP/mlquestions
Updated • 194 • 3 -
McGill-NLP/bart-qg-mlquestions-backtraining
Updated • 8 -
McGill-NLP/bart-qg-mlquestions-selftraining
Updated • 5
Best open African LLM
-
AfriqueLLM: How Data Mixing and Model Architecture Impact Continued Pre-training for African Languages
Paper • 2601.06395 • Published • 5 -
McGill-NLP/AfriqueQwen-14B
Text Generation • 15B • Updated • 2.32k • • 4 -
McGill-NLP/AfriqueQwen-8B
Text Generation • 8B • Updated • 1.58k • • 2 -
McGill-NLP/AfriqueQwen3.5-4B-50Langs
Text Generation • 5B • Updated • 399 • 6
Models and data from "Structured Distillation of Web Agent Capabilities Enables Generalization" (arXiv:2604.07776)
-
Structured Distillation of Web Agent Capabilities Enables Generalization
Paper • 2604.07776 • Published • 23 -
McGill-NLP/A3-Qwen3.5-9B
Image-Text-to-Text • 9B • Updated • 283 • 6 -
McGill-NLP/A3-Qwen3.5-4B
Image-Text-to-Text • 5B • Updated • 91 • 2 -
McGill-NLP/A3-Qwen3.5-2B
Image-Text-to-Text • 3B • Updated • 31 • 2
Generative Embeddings from Large Language Models
Pre-computed contextual text embeddings for interpreting LLM/VLM hidden states. Use with: pip install latentlens
Reformulating the RL of reasoning LLMs through Markovian Thinking paradigm.
INJONGO: A Multicultural Intent Detection and Slot-filling Dataset for 16 African Languages
-
McGill-NLP/AfroXLMR-large-76L-Injongo-intent
Text Classification • 0.6B • Updated • 3 -
McGill-NLP/AfroXLMR-large-76L-Injongo-slot
Token Classification • 0.6B • Updated • 4 -
McGill-NLP/gemma-2-9b-it-Injongo-intent
Text Generation • 9B • Updated • 3 -
McGill-NLP/gemma-2-9b-it-Injongo-slot
Text Generation • 9B • Updated • 2
Datasets used for the OLMo experiments in the "Not All Data are Unlearned Equally" paper https://arxiv.org/abs/2504.05058
-
AgentRewardBench: Evaluating Automatic Evaluations of Web Agent Trajectories
Paper • 2504.08942 • Published • 29 -
McGill-NLP/agent-reward-bench
Viewer • Updated • 1.41k • 7.35k • 4 -
Agent Reward Bench Demo
💻5Explore agent trajectories and judgments in web benchmarks
-
Agent Reward Bench Leaderboard
🥇3Leaderboard for AgentRewardBench
Generate challenging synthetic data to evaluate LLMs
-
McGill-NLP/LLM2Vec-Meta-Llama-32-3B-Instruct-mntp-supervised
Updated -
McGill-NLP/LLM2Vec-Meta-Llama-31-8B-Instruct-mntp-supervised
Sentence Similarity • Updated • 317 • 5 -
McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised
Sentence Similarity • Updated • 112k • 52 -
McGill-NLP/LLM2Vec-Mistral-7B-Instruct-v2-mntp-supervised
Sentence Similarity • Updated • 270 • 13
https://mcgill-nlp.github.io/weblinx
Repository: https://github.com/McGill-NLP/AURORA
https://mcgill-nlp.github.io/weblinx
mcgill-nlp.github.io/statcan-dialogue-dataset
-
The StatCan Dialogue Dataset: Retrieving Data Tables through Conversations with Genuine Intents
Paper • 2304.01412 • Published • 2 -
McGill-NLP/statcan-dialogue-dataset
Preview • Updated • 4 • 7 -
McGill-NLP/dpr-statcan-conversation_encoder-title
Feature Extraction • 0.1B • Updated • 7 -
McGill-NLP/tapas-statcan-large-conversation_encoder-cell_tokens
Feature Extraction • Updated • 4
-
Back-Training excels Self-Training at Unsupervised Domain Adaptation of Question Generation and Passage Retrieval
Paper • 2104.08801 • Published • 1 -
McGill-NLP/mlquestions
Updated • 194 • 3 -
McGill-NLP/bart-qg-mlquestions-backtraining
Updated • 8 -
McGill-NLP/bart-qg-mlquestions-selftraining
Updated • 5