Update README.md

README.md

@@ -1,3 +1,48 @@
----
-license: mit
----
+---
+license: mit
+language:
+- en
+base_model:
+- google/gemma-3-1b-it
+---
+# Emma-3-1B: Emotionally Modulated Gemma-3
+
+**Model Author:** FelixTheWhale
+
+## Model Description
+
+**Emma-3-1B** (Emotional Gemma 3) is an experimental implementation exploring emotional modulation within the Gemma-3 LLM architecture. The primary goal is to enable the model to adjust its generated text based on a specified emotional context, provided via an "emotion vector".
+
+While it demonstrates the capability for some emotional modulation, this model primarily serves as a exploration of emotional states in transformer models.
+
+### Emotion Representation
+
+**8 emotion dimensions**:
+
+*   SADNESS ↔ JOY
+*   FEAR ↔ COURAGE
+*   DISGUST ↔ ACCEPTANCE
+*   ANGER ↔ CALMNESS
+*   SURPRISE ↔ EXPECTATION
+*   DISTRUST ↔ TRUST
+*   BOREDOM ↔ INTEREST
+*   INDIFFERENCE ↔ EMPATHY
+
+Each dimension is represented by a value (e.g., between -1 and 1), forming an 8-dimensional vector input.
+
+## How it Works: Architecture
+
+1.  **Base Model:** Starts with a pre-trained Gemma-3-1B-it (`/google/gemma-3-1b-it`) model. Also may work with other models with adjustments in forward().
+2.  **Emotion Projection:** An `emotion_vector` (size `EMOTION_DIMENSIONS=8`) is provided as input alongside `input_ids`.
+3.  **Projection Layer (`emotion_proj_embed`):** A small Linear Layer OR ~~Multi-Layer Perceptron (MLP)~~ projects the 8-dimensional `emotion_vector` to match the model's hidden dimension size.
+4.  **Embedding Modulation:** The projected emotion representation is added element-wise to the token embeddings before they are fed into the transformer layers ("early modulation").
+5.  **Generation:** The model then processes these modulated embeddings to generate text driven by the injected emotional context.
+
+*(Note: The model class inherits from `transformers.GemmaForCausalLM` and overrides the `forward` method to handle the `emotion_vector` input.)*
+
+## Training
+
+*   **Fine-tuning:** The model was fine-tuned using Parameter-Efficient Fine-Tuning (PEFT), specifically LoRA (Low-Rank Adaptation). Only the LORA adapters and the `emotion_proj_embed` layer were trained.
+*   **Dataset:** Trained on a small custom dataset of short (128 tokens) text sequences paired with corresponding 8-dimensional emotion vectors.
+*   **Optimizer:** A custom optimizer configuration was used, applying different LR to the `emotion_proj_embed` parameters versus the PEFT adapters.
+*   **Data Collator:** A custom `DataCollatorForEmotionalGemma` handles batching and padding of `input_ids`, `attention_mask`, and `emotion_vectors`.