Legal-verse/sinergi-model-evaluation-results

Sinergi model evaluation outputs

Model answers and generation telemetry used by the Sinergi Table 8-aligned evaluation notebook. The repository contains eight configurations so every system can be loaded independently with the Hugging Face datasets library.

from datasets import load_dataset

data = load_dataset(
    "Legal-verse/sinergi-model-evaluation-results",
    "qwen-sft-rl-rag",
    split="test",
)

Configurations

Configuration	System	Rows	Source file
deepseek-sft	DeepSeek-R1-Distill-Qwen-1.5B SFT	9,176	deepseek-r1-distill-qwen-1-5b-test-outputs.parquet
deepseek-sft-rag	DeepSeek-R1-Distill-Qwen-1.5B SFT+RAG	2,108	deepseek-r1-distill-qwen-1-5b-rag-test-outputs.parquet
gemma-sft	Gemma-4-E2B SFT	9,176	gemma-4-e2b-test-outputs.parquet
gemma-sft-rag	Gemma-4-E2B SFT+RAG	2,108	gemma-4-e2b-rag-test-outputs.parquet
qwen-sft	Qwen3.5-4B SFT	9,176	qwen3-5-4b-test-outputs.parquet
qwen-sft-rag	Qwen3.5-4B SFT+RAG	2,108	qwen3-5-4b-rag-test-outputs.parquet
qwen-sft-rl	Qwen3.5-4B SFT+RL	9,176	qwen3-5-4b-test-outputs rl.parquet
qwen-sft-rl-rag	Qwen3.5-4B SFT+RL+RAG	2,108	qwen3-5-4b-rag-test-outputs rl.parquet

Data scope

Each row contains a model answer, its gold structured answer, identifiers/provenance, generation token counts, finish state, and chunk-level timing. RAG configurations additionally contain retrieval references, scores, and grounding diagnostics. The SFT and RAG configurations are different held-out datasets and must not be treated as paired examples.

The source research paper is not included in this dataset repository. Evaluation notebooks, computed metric tables, figures, and run metadata are backed up separately to the existing W&B project.