Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,97 @@
+---
+tags:
+  - humor
+  - computational-humor
+  - cognitive-synergy-framework
+  - collection
+license: apache-2.0
+---
+
+# HumorGen
+
+**SaLT Lab, Carnegie Mellon University**
+
+---
+
+## Overview
+
+Training language models to generate high-quality humor is fundamentally difficult.
+Standard next-token prediction pushes models toward the most probable — and therefore
+the safest, most generic — continuation. In humor, that failure mode is well-known:
+models produce predictable, unfunny outputs.
+
+HumorGen addresses this through the **Cognitive Synergy Framework (CSF)**, a
+Mixture-of-Thought approach in which a teacher model generates humor from six distinct
+cognitive personas, each grounded in a psychological theory of comedy. The resulting
+candidate pool is diverse by construction. A compact student model (7B) is then
+trained on this data, learning that a single input can support multiple valid comedic
+interpretations. Preference alignment via DPO or Offline GRPO further refines which
+comedic angle the model selects in context.
+
+Full methodology: [arXiv:2604.09629](https://arxiv.org/abs/2604.09629)
+
+---
+
+## Constrained Humor: HumorGen at CLEF 2026 JOKER
+
+The CSF generalizes naturally to constrained humor generation. The CLEF 2026 JOKER
+Task 4 requires producing a pun-brief — a sentence that simultaneously satisfies a
+given pun word and two specified semantic senses. This is a significantly harder
+setting than open-ended headline humor: the model must navigate strict lexical
+constraints while still producing output that reads as genuinely funny.
+
+To scale the framework to this multilingual, constrained task, we first train
+domain-agnostic humor pretraining checkpoints at 14B and 32B scale (HumorGen_SFT_14B
+and HumorGen_SFT_32B) on the full SemEval MWAHAHA corpus across all languages.
+These serve as strong multilingual starting points and are released independently as
+general-purpose multilingual humor generators. The JOKER-specific models are then
+branched from these bases via per-language LoRA fine-tuning.
+
+Full methodology: [HumorGen at CLEF 2026 JOKER Task 4](https://edwardajayi.github.io/assets/papers/HumorGen-JOKER.pdf)
+
+---
+
+## Models
+
+| Model | Description | Backbone |
+|:---|:---|:---|
+| [HumorGen_SFT_7B](https://huggingface.co/Jayi2424/HumorGen_SFT_7B) | SFT baseline | Qwen2.5-7B |
+| [HumorGen_SFT_Think_7B](https://huggingface.co/Jayi2424/HumorGen_SFT_Think_7B) | SFT + Chain-of-Thought | Qwen2.5-7B |
+| [HumorGen_DPO_7B](https://huggingface.co/Jayi2424/HumorGen_DPO_7B) | DPO | Qwen2.5-7B |
+| [HumorGen_DPO_Think_7B](https://huggingface.co/Jayi2424/HumorGen_DPO_Think_7B) | DPO + Chain-of-Thought | Qwen2.5-7B |
+| [HumorGen_GRPO_7B](https://huggingface.co/Jayi2424/HumorGen_GRPO_7B) | O-GRPO | Qwen2.5-7B |
+| [HumorGen_GRPO_Think_7B](https://huggingface.co/Jayi2424/HumorGen_GRPO_Think_7B) | O-GRPO + Chain-of-Thought | Qwen2.5-7B |
+| [HumorGen_SFT_14B](https://huggingface.co/Jayi2424/HumorGen_SFT_14B) | Multilingual SFT base (used in JOKER) | Qwen3-14B |
+| [HumorGen_SFT_32B](https://huggingface.co/Jayi2424/HumorGen_SFT_32B) | Multilingual SFT base (used in JOKER) | Qwen3-32B |
+| [HumorGen_JOKER_EN_14B](https://huggingface.co/Jayi2424/HumorGen_JOKER_EN_14B) | CLEF JOKER Task 4 — English, 14B | Qwen3-14B |
+| [HumorGen_JOKER_EN_32B](https://huggingface.co/Jayi2424/HumorGen_JOKER_EN_32B) | CLEF JOKER Task 4 — English, 32B | Qwen3-32B |
+| [HumorGen_JOKER_FR_14B](https://huggingface.co/Jayi2424/HumorGen_JOKER_FR_14B) | CLEF JOKER Task 4 — French, 14B | Qwen3-14B |
+| [HumorGen_JOKER_FR_32B](https://huggingface.co/Jayi2424/HumorGen_JOKER_FR_32B) | CLEF JOKER Task 4 — French, 32B | Qwen3-32B |
+| [HumorGen_JOKER_ES_14B](https://huggingface.co/Jayi2424/HumorGen_JOKER_ES_14B) | CLEF JOKER Task 4 — Spanish, 14B | Qwen3-14B |
+| [HumorGen_JOKER_ES_32B](https://huggingface.co/Jayi2424/HumorGen_JOKER_ES_32B) | CLEF JOKER Task 4 — Spanish, 32B | Qwen3-32B |
+
+---
+
+## Citation
+
+```bibtex
+@misc{ajayi2026humorgen,
+  title         = {HumorGen: Cognitive Synergy for Humor Generation in Large Language
+                   Models via Persona-Based Distillation},
+  author        = {Ajayi, Edward and others},
+  year          = {2026},
+  eprint        = {2604.09629},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.CL},
+  url           = {https://arxiv.org/abs/2604.09629}
+}
+
+@inproceedings{ajayi2026joker,
+  title     = {HumorGen at CLEF 2026 JOKER Task 4: Cross-Lingual Constrained
+               Pun Generation via the Cognitive Synergy Framework},
+  author    = {Ajayi, Edward and others},
+  booktitle = {Working Notes of CLEF 2026},
+  year      = {2026},
+  url       = {https://edwardajayi.github.io/assets/papers/HumorGen-JOKER.pdf}
+}
+```