Update README.md

README.md

@@ -9,7 +9,9 @@ tags:
 - commit-message-generation
 - code-summarization
 - generated_from_trainer
-license: apache-2.0
+license: cc-by-nc-4.0
+datasets:
+- Maxscha/commitbench
 language:
 - en
 ---
@@ -33,7 +35,8 @@ This model is a **QLoRA (4-bit quantized LoRA)** adapter trained on the Qwen2.5-
 - **Developed by:** Mamoun Yosef
 - **Model type:** Causal Language Model (Decoder-only Transformer) with LoRA adapters
 - **Language(s):** English
-- **License:** Apache 2.0
+- **License:** CC BY-NC 4.0 (non-commercial for this trained adapter)
+- **Base model license:** Apache 2.0 (`Qwen/Qwen2.5-Coder-0.5B`)
 - **Finetuned from model:** Qwen/Qwen2.5-Coder-0.5B
 
 ### Model Sources
@@ -41,6 +44,12 @@ This model is a **QLoRA (4-bit quantized LoRA)** adapter trained on the Qwen2.5-
 - **Repository:** [commit-message-llm](https://github.com/mamounyosef/commit-message-llm)
 - **Base Model:** [Qwen/Qwen2.5-Coder-0.5B](https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B)
 
+## License and Usage
+
+- This adapter was trained using **CommitBench** (`Maxscha/commitbench`), licensed **CC BY-NC 4.0**.
+- This trained adapter is therefore **non-commercial use only**.
+- The base model (`Qwen/Qwen2.5-Coder-0.5B`) remains licensed under **Apache-2.0**.
+
 ## Uses
 
 ### Direct Use
@@ -88,6 +97,7 @@ Can be integrated into:
 - Diffs from non-programming languages
 - Extremely large diffs (>8000 characters)
 - Commit messages requiring deep domain knowledge beyond code structure
+- Commercial usage of this trained adapter
 
 ## Bias, Risks, and Limitations
 
@@ -174,9 +184,9 @@ print(message)
 **Preprocessing:**
 - Removed trivial messages (fix, update, wip, etc.)
 - Filtered out reference-only commits (fix #123)
-- Removed placeholder tokens (<HASH>, <URL>)
+- Removed placeholder tokens (`<HASH>`, `<URL>`)
 - Kept diffs between 50-8000 characters
-- Required messages with semantic content (≥3 words)
+- Required messages with semantic content (>=3 words)
 
 **Final dataset sizes:**
 - Training: 120,000 samples
@@ -197,10 +207,10 @@ Prompt tokens (diff + separator) are masked with label `-100` so loss is compute
 
 #### Preprocessing
 
-1. Normalize newlines (CRLF → LF)
+1. Normalize newlines (CRLF -> LF)
 2. Tokenize diff + separator + message
 3. Mask prompt labels to `-100`
-4. Truncate to max_length=512 tokens
+4. Truncate to `max_length=512` tokens
 5. Append EOS token to target
 
 #### Training Hyperparameters
@@ -257,7 +267,7 @@ Prompt tokens (diff + separator) are masked with label `-100` so loss is compute
 - **Loss:** Cross-entropy loss on commit message tokens
 - **Perplexity:** exp(loss), measures model confidence
   - Lower perplexity = better prediction quality
-  - Perplexity ≈ 17 is strong for this task
+  - Perplexity ~17 is strong for this task
 
 ### Results