Upload folder using huggingface_hub

README.md

@@ -18,72 +18,68 @@ size_categories:
 
 # Open Telco Leaderboard Dataset
 
-This dataset contains evaluation results from the **Open Telco** benchmark suite, which measures LLM performance on telecommunications-specific tasks.
+Raw benchmark scores for LLMs evaluated on telecommunications-specific tasks. This minimal dataset contains only essential fields - all derived metrics (rank, mean, TCI) should be calculated from these scores.
 
-## Dataset Description
-
-The leaderboard tracks model performance across four specialized telecom benchmarks:
-
-| Benchmark | Description | Samples |
-|-----------|-------------|---------|
-| **TeleQnA** | 10,000 Q&A pairs testing telecom knowledge across lexicon, research, and standards | 10,000 |
-| **TeleMath** | Mathematical reasoning in telecommunications (signal processing, network optimization) | 500 |
-| **TeleLogs** | Root cause analysis for 5G network throughput degradation across 8 failure modes | 1,000+ |
-| **3GPP-TSG** | Classification of technical documents by 3GPP working group (RAN, SA, CT) | 5,000+ |
-
-## Data Schema
-
-The parquet file contains the following columns:
+## Schema
 
 | Column | Type | Description |
 |--------|------|-------------|
-| `rank` | int | Model's overall ranking position |
 | `model` | string | Model name (e.g., "gpt-5.2", "claude-opus-4.5") |
-| `provider` | string | Provider name (e.g., "OpenAI", "Anthropic") |
-| `repo` | string | Full model path |
-| `mean` | float | Mean score across all benchmarks |
-| `teleqna` | float | TeleQnA benchmark score |
-| `teleqna_stderr` | float | TeleQnA standard error |
-| `telelogs` | float | TeleLogs benchmark score |
-| `telelogs_stderr` | float | TeleLogs standard error |
-| `telemath` | float | TeleMath benchmark score |
-| `telemath_stderr` | float | TeleMath standard error |
-| `tsg` | float | 3GPP-TSG benchmark score |
-| `tsg_stderr` | float | 3GPP-TSG standard error |
+| `provider` | string | Provider (e.g., "OpenAI", "Anthropic") |
+| `teleqna` | float | TeleQnA benchmark score (0-100) |
+| `teleqna_stderr` | float | Standard error |
+| `telelogs` | float | TeleLogs benchmark score (0-100) |
+| `telelogs_stderr` | float | Standard error |
+| `telemath` | float | TeleMath benchmark score (0-100) |
+| `telemath_stderr` | float | Standard error |
+| `tsg` | float | 3GPP-TSG benchmark score (0-100) |
+| `tsg_stderr` | float | Standard error |
+
+## Benchmarks
+
+| Benchmark | Description | Samples |
+|-----------|-------------|---------|
+| **TeleQnA** | Q&A pairs testing telecom knowledge | 10,000 |
+| **TeleMath** | Mathematical reasoning in telecommunications | 500 |
+| **TeleLogs** | Root cause analysis for 5G network issues | 1,000+ |
+| **3GPP-TSG** | Classification of 3GPP technical documents | 5,000+ |
 
 ## Usage
 
 ```python
 from datasets import load_dataset
+import pandas as pd
+
+# Load dataset
+ds = load_dataset("GSMA/leaderboard")
+df = ds["train"].to_pandas()
 
-dataset = load_dataset("GSMA/leaderboard")
-df = dataset["train"].to_pandas()
+# Calculate mean score
+benchmarks = ['teleqna', 'telelogs', 'telemath', 'tsg']
+df['mean'] = df[benchmarks].mean(axis=1)
 
-# View top models
-print(df.sort_values("mean", ascending=False).head(10))
+# Rank by mean score
+df['rank'] = df['mean'].rank(ascending=False).astype(int)
+
+print(df.sort_values('rank'))
 ```
 
-## How Results Are Generated
+## Derived Metrics
 
-Evaluations are run using the [Inspect AI](https://inspect.ai-safety-institute.org.uk/) framework. The workflow:
+These fields are NOT stored but can be calculated:
 
-1. Models are evaluated against each benchmark
-2. Results are logged as `.eval` files
-3. Scores are aggregated into this parquet file
-4. The leaderboard website displays the rankings
+- **rank**: Sort by mean or TCI score
+- **mean**: Average of the 4 benchmark scores
+- **TCI (Telco Capability Index)**: IRT-inspired score using benchmark difficulties
 
 ## Related Datasets
 
-- [TeleQnA](https://huggingface.co/datasets/netop/TeleQnA) - Telecom Q&A benchmark
-- [TeleMath](https://huggingface.co/datasets/netop/TeleMath) - Telecom math reasoning
-- [TeleLogs](https://huggingface.co/datasets/netop/TeleLogs) - Network log analysis
-- [3GPP-TSG](https://huggingface.co/datasets/eaguaida/gsma_sample) - 3GPP document classification
+- [TeleQnA](https://huggingface.co/datasets/netop/TeleQnA)
+- [TeleMath](https://huggingface.co/datasets/netop/TeleMath)
+- [TeleLogs](https://huggingface.co/datasets/netop/TeleLogs)
+- [3GPP-TSG](https://huggingface.co/datasets/eaguaida/gsma_sample)
 
 ## Links
 
 - [Open Telco Website](https://gsma-research.github.io/open_telco/)
-- [GitHub Repository](https://github.com/gsma-research/open_telco)
-
-## License
-
-Apache 2.0
+- [GitHub](https://github.com/gsma-research/open_telco)

evals.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:56db654b1ad079f16a8c59a44806d918005cec7c8738a939f34b250df4a63283
-size 47503
+oid sha256:e1d174838cb72b1bb7ac747a9df6e5eff23ea8ad3c9c2c93947731a9230278b9
+size 6608