Updated Readme

README.md

@@ -41,7 +41,7 @@ dataset_info:
 
 Dataset for the paper **"Beyond Accuracy: Characterizing Code Comprehension Capabilities in (Large) Language Models"** by Machtle, Serr, Loose & Eisenbarth (University of Luebeck).
 
-[[Paper]](https://arxiv.org/abs/2601.12951) | [[Code]](https://github.com/Felix6326727/beyond-accuracy-code-comprehension)
+[[Paper]](https://arxiv.org/abs/2601.12951) | [[Code]](https://github.com/UzL-ITS/code-comprehension-capabilities-llms/)
 
 ## Task
 
@@ -68,7 +68,6 @@ print(f"Cyclomatic complexity: {sample['metric_cyclomatic_complexity']}")
 
 | | |
 |---|---|
-| **Samples** | 12,584 |
 | **Columns** | 249 |
 | **Source** | Python subset of [Project CodeNet](https://github.com/IBM/Project_CodeNet) |
 | **I/O generation** | Type-aware fuzzing with hill-climbing type inference |
@@ -102,15 +101,6 @@ Per-model results from the binary I/O consistency evaluation. Each model has 3 c
 | `llm_{model}_num_correct` | int | Number of test cases answered correctly (out of `num_total`) |
 | `llm_{model}_num_total` | int | Total test cases for this sample (typically 2: one correct, one incorrect) |
 
-**Models:**
-
-| Column prefix | Model | Samples evaluated |
-|---|---|---|
-| `llm_gpt_oss_120b_` | GPT-OSS 120B | 12,509 |
-| `llm_llama_3_3_70b_` | Llama 3.3 70B Instruct | 12,517 |
-| `llm_mistral_small_24b_` | Mistral Small 24B Instruct | 12,301 |
-| `llm_phi4_` | Phi-4 | 12,596 |
-| `llm_codellama_13b_` | CodeLlama 13B Instruct | 3,470 |
 
 ### Code Metric Columns (224)
 
@@ -122,50 +112,6 @@ All prefixed with `metric_`. Values are floats (or null if unavailable).
 
 **Opcode Statistics (39 columns)** — Python bytecode features: `num_opcodes`, `sum_opcodes`, `avg_opcode_count`, `min_opcode_count`, `max_opcode_count`, individual opcode counts (`opcode_1`, `opcode_83`, ...), `opcodes_used0`–`opcodes_used3`, and `top_0_opcode_name` through `top_19_opcode_name`.
 
-## Key Results from the Paper
-
-### LLM Performance on the I/O Consistency Task
-
-| Model | Accuracy | F1 |
-|---|---|---|
-| GPT-OSS 120B | 0.960 | 0.959 |
-| Mistral Small 24B Instruct | 0.744 | 0.685 |
-| Llama 3.3 70B Instruct | 0.738 | 0.662 |
-| Phi-4 | 0.733 | 0.674 |
-| CodeLlama 13B Instruct | 0.506 | 0.062 |
-
-### Do Human Metrics Predict LLM Success?
-
-A classifier trained on the 224 code metrics achieves a mean **AUROC of 0.63** across models — weak correlation. A *shadow model* (fine-tuned UniXcoder on raw code + I/O) reaches **AUROC 0.86**, confirming LLMs use patterns not captured by traditional software engineering metrics.
-
-## Example Usage
-
-### Filter samples where GPT-OSS succeeded but Phi-4 failed
-
-```python
-ds_filtered = ds.filter(
-    lambda x: x["llm_gpt_oss_120b_success"] == True
-              and x["llm_phi4_success"] == False
-)
-print(f"{len(ds_filtered)} samples where GPT-OSS succeeded but Phi-4 failed")
-```
-
-### Analyze metrics vs. model success
-
-```python
-import pandas as pd
-
-df = ds.to_pandas()
-metric_cols = [c for c in df.columns if c.startswith("metric_") and df[c].dtype == "float64"]
-
-# Compare mean complexity for GPT-OSS successes vs failures
-success = df[df["llm_gpt_oss_120b_success"] == True][metric_cols].mean()
-failure = df[df["llm_gpt_oss_120b_success"] == False][metric_cols].mean()
-
-diff = (failure - success).sort_values(ascending=False)
-print("Metrics most elevated in GPT-OSS failures:")
-print(diff.head(10))
-```
 
 ## Data Generation Pipeline
 
@@ -185,7 +131,7 @@ Python files (CodeNet)
   This dataset
 ```
 
-See the [GitHub repository](https://github.com/Felix6326727/beyond-accuracy-code-comprehension) for the full pipeline code.
+See the [GitHub repository](https://github.com/UzL-ITS/code-comprehension-capabilities-llms/) for the full pipeline code.
 
 ## Citation