sdg-hub 0.2.2__tar.gz → 0.3.0__tar.gz

{sdg_hub-0.2.2 → sdg_hub-0.3.0}/.github/workflows/pypi.yaml

@@ -78,7 +78,7 @@ jobs:
                   path: dist
 
             - name: "Upload to Test PyPI"
-              uses: pypa/gh-action-pypi-publish@76f52bc884231f62b9a034ebfe128415bbaabdfc # v1.12.4
+              uses: pypa/gh-action-pypi-publish@ed0c53931b1dc9bd32cbe73a98c7f6766f8a527e # v1.13.0
               with:
                   repository-url: https://test.pypi.org/legacy/
 
@@ -130,4 +130,4 @@ jobs:
                   rm ./dist/*.sigstore.json
 
             - name: "Upload to PyPI"
-              uses: pypa/gh-action-pypi-publish@76f52bc884231f62b9a034ebfe128415bbaabdfc # v1.12.4
+              uses: pypa/gh-action-pypi-publish@ed0c53931b1dc9bd32cbe73a98c7f6766f8a527e # v1.13.0

{sdg_hub-0.2.2/src/sdg_hub.egg-info → sdg_hub-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sdg_hub
-Version: 0.2.2
+Version: 0.3.0
 Summary: Synthetic Data Generation
 Author-email: Red Hat AI Innovation <abhandwa@redhat.com>
 License: Apache-2.0
@@ -53,6 +53,7 @@ Requires-Dist: sentence-transformers; extra == "examples"
 Requires-Dist: instructor; extra == "examples"
 Requires-Dist: fastapi; extra == "examples"
 Requires-Dist: nest-asyncio; extra == "examples"
+Requires-Dist: ipykernel; extra == "examples"
 Provides-Extra: dev
 Requires-Dist: pre-commit<4.0,>=3.0.4; extra == "dev"
 Requires-Dist: pylint<4.0,>=2.16.2; extra == "dev"
@@ -63,6 +64,7 @@ Requires-Dist: pytest-cov; extra == "dev"
 Requires-Dist: pytest-html; extra == "dev"
 Requires-Dist: tox<5,>=4.4.2; extra == "dev"
 Requires-Dist: ruff; extra == "dev"
+Requires-Dist: pytest-env; extra == "dev"
 Dynamic: license-file
 
 # `sdg_hub`: Synthetic Data Generation Toolkit

sdg_hub-0.3.0/examples/knowledge_tuning/enhanced_summary_knowledge_tuning/.env.example

@@ -0,0 +1,59 @@
+# SDG Knowledge Tuning Configuration
+# Copy this file to .env and update the values as needed
+
+# =============================================================================
+# MODEL CONFIGURATION
+# Choose one of the following model providers: hosted_vllm, openai, ollama
+# =============================================================================
+
+MODEL_PROVIDER=hosted_vllm
+LITELLM_MODE=PRODUCTION
+
+# =============================================================================
+# HOSTED VLLM CONFIGURATION
+# =============================================================================
+VLLM_MODEL=hosted_vllm/meta-llama/Llama-3.3-70B-Instruct
+VLLM_API_BASE=http://localhost:8000/v1
+VLLM_API_KEY=EMPTY
+ENABLE_REASONING=false
+# =============================================================================
+# OPENAI CONFIGURATION
+# =============================================================================
+OPENAI_API_KEY=your-openai-api-key-here  
+OPENAI_MODEL=openai/gpt-5  
+
+# =============================================================================
+# OLLAMA CONFIGURATION
+# =============================================================================
+OLLAMA_MODEL=ollama/gemma3
+OLLAMA_API_BASE=http://localhost:11434
+
+# =============================================================================
+# MAAS CONFIGURATION (Red Hat AI Services)
+# =============================================================================
+MAAS_MODEL=your-provisioned-model-name
+MAAS_API_BASE=your-provisioned-model-url
+MAAS_API_KEY=your-maas-api-key-here
+
+# =============================================================================
+# DATA CONFIGURATION
+# =============================================================================
+SEED_DATA_PATH=seed_data_val.jsonl
+OUTPUT_DATA_FOLDER=output_data
+RUN_ON_VALIDATION_SET=true
+NUMBER_OF_SUMMARIES=50
+
+# =============================================================================
+# DATA MIXING CONFIGURATION
+# =============================================================================
+SAVE_GPT_OSS_FORMAT=false
+STUDENT_MODEL=meta-llama/Llama-3.1-8B-Instruct
+
+# Cut sizes for data mixing (comma-separated list of integers)
+# Number of summaries to pick per document for each cut
+CUT_SIZES=3,5,7
+
+# Number of Q&A pairs per document
+QA_PER_DOC=3
+
+HF_TOKEN=your-hf-token

sdg_hub-0.3.0/examples/knowledge_tuning/enhanced_summary_knowledge_tuning/README.md

@@ -0,0 +1,76 @@
+# Knowledge Tuning with Enhanced Summaries
+
+## Objective
+
+Pre-trained language models typically encounter most facts in their training data only **once or twice**, if at all. As a result, knowledge of specific details—especially **proprietary or domain-specific documents**—is often incomplete or missing.
+
+This pipeline is designed to **inject new knowledge** from a given set of documents into an instruction-tuned model. By generating **multiple document augmentations** (summaries, extractive passages, atomic facts) and **synthetic Q\&A pairs**, we repeat and reinforce important information. This repetition helps the model:
+
+* **Memorize facts** it has rarely or never seen before.
+* **Generalize across augmentations**, improving reliability when queried.
+* **Adapt to proprietary knowledge sources** that were absent from pre-training.
+
+The final product is a **high-quality training dataset** suitable for fine-tuning, enabling models to answer queries more accurately and faithfully based on the injected documents.
+
+---
+
+## 1. Document Summarization
+
+To bootstrap the process, we generate **three complementary types of summaries** for each source document. This ensures the model captures content at multiple levels of abstraction:
+
+* **Detailed Summaries** – Rich, comprehensive overviews of the document.
+* **Extractive Summaries** – Directly extracted sentences and passages representing the most important parts.
+* **Atomic Facts** – Concise, standalone factual statements distilled from the text.
+
+This multi-perspective approach improves the model’s ability to **memorize, generalize, and recall** key knowledge.
+
+---
+
+## 2. Synthetic Q\&A Generation
+
+With summaries in place, we scale up training data via **synthetic Q\&A generation**:
+
+* Users provide a small set of **seed examples** (initial Q\&A pairs).
+* The pipeline uses these seeds to generate a large set of **contextually grounded Q\&A pairs**, tightly linked to the summarized documents.
+* This expands sparse seed data into a **rich, diverse training dataset** suitable for fine-tuning.
+
+---
+
+## 3. Quality Control
+
+High-quality training data is essential. To ensure faithfulness and accuracy, we employ a **teacher-model evaluation loop**:
+
+1. Provide the model with a generated answer and the original document.
+2. Ask it to extract each factual claim from the answer.
+3. Verify whether each claim is **explicitly supported** by the document.
+
+Only claims passing this check are retained. This process filters out **hallucinations and unsupported statements**, ensuring reliable Q\&A pairs.
+
+---
+
+## Data Generation Statistics
+
+### Summary Augmentation
+
+Each “cut” represents the total number of summaries generated per document across all three augmentation types.
+
+| Cut (NUMBER\_OF\_SUMMARIES = 3) | Token Count |
+| ------------------------------- | ----------- |
+| 1                               | 2,193,502   |
+| 2                               | 4,383,655   |
+| 5                               | 10,870,396  |
+| 10                              | 21,815,170  |
+| 20                              | 43,601,976  |
+| 30                              | 65,395,710  |
+| 40                              | 87,118,308  |
+| 50                              | 108,779,213 |
+
+---
+
+### Finance Bench Example
+
+For Finance Bench (NUMBER\_OF\_SUMMARIES = 1):
+
+| Cut | Token Count |
+| --- | ----------- |
+| 50  | 213,333,192 |