Translated ['src/AI/AI-llm-architecture/2.-data-sampling.md'] to de

src/AI/AI-llm-architecture/2.-data-sampling.md

@@ -230,9 +230,70 @@ tensor([[  367,  2885,  1464,  1807],
 [ 3285,   326,    11,   287]])
 ]
 ```
-## Referenzen
+## Fortgeschrittene Sampling-Strategien (2023-2025)
 
-- [https://www.manning.com/books/build-a-large-language-model-from-scratch](https://www.manning.com/books/build-a-large-language-model-from-scratch)
+### 1. Temperaturbasierte Mischgewichtung
+State-of-the-art LLMs werden selten auf einem einzelnen Korpus trainiert. Stattdessen sampeln sie aus mehreren heterogenen Datenquellen (Code, Web, wissenschaftliche Arbeiten, Foren…). Der relative Anteil jeder Quelle kann die nachgelagerte Leistung stark beeinflussen. Neuere Open-Source-Modelle wie Llama 2 führten ein **temperaturbasiertes Sampling-Schema** ein, bei dem die Wahrscheinlichkeit, ein Dokument aus dem Korpus *i* zu ziehen, wird
+```
+p(i) = \frac{w_i^{\alpha}}{\sum_j w_j^{\alpha}}
+```
+• *w<sub>i</sub>* – Rohtoken-Prozentsatz des Korpus *i*  
+• *α* ("Temperatur") – ein Wert in (0,1].  α < 1 flacht die Verteilung ab und gewichtet kleinere, qualitativ hochwertige Korpora stärker.
 
+Llama 2 verwendete α = 0.7 und zeigte, dass eine Verringerung von α die Bewertungsergebnisse bei wissensintensiven Aufgaben verbesserte, während die Trainingsmischung stabil blieb. Der gleiche Trick wird von Mistral (2023) und Claude 3 übernommen.
+```python
+from collections import Counter
+
+def temperature_sample(corpus_ids, alpha=0.7):
+counts = Counter(corpus_ids)           # number of tokens seen per corpus
+probs  = {c: c_count**alpha for c, c_count in counts.items()}
+Z = sum(probs.values())
+probs = {c: p/Z for c, p in probs.items()}
+# Now draw according to probs to fill every batch
+```
+
+```
+
+### 2. Sequence Packing / Dynamic Batching
+GPU memory is wasted when every sequence in a batch is padded to the longest example.  "Packing" concatenates multiple shorter sequences until the **exact** `max_length` is reached and builds a parallel `attention_mask` so that tokens do not attend across segment boundaries.  Packing can improve throughput by 20–40 % with no gradient change and is supported out-of-the-box in
+
+* PyTorch `torchtext.experimental.agents.PackedBatch`
+* HuggingFace `DataCollatorForLanguageModeling(pad_to_multiple_of=…)`
+
+Dynamic batching frameworks (e.g. FlashAttention 2, vLLM 2024) combine sequence packing with just-in-time kernel selection, enabling thousand-token context training at 400+ K tokens/s on A100-80G.
+
+### 3. Deduplication & Quality Filtering
+Repeated passages cause memorization and provide an easy channel for data-poisoning.  Modern pipelines therefore:
+
+1. MinHash/FAISS near-duplicate detection at **document** and **128-gram** level.
+2. Filter documents whose perplexity under a small reference model is > µ + 3σ (noisy OCR, garbled HTML).
+3. Block-list documents that contain PII or CWE keywords using regex & spaCy NER.
+
+The Llama 2 team deduplicated with 8-gram MinHash and removed ~15 % of CommonCrawl before sampling.  OpenAI’s 2024 "Deduplicate Everything" paper demonstrates ≤0.04 duplicate ratio reduces over-fitting and speeds convergence.
+
+## Security & Privacy Considerations During Sampling
+
+### Data-Poisoning / Backdoor Attacks
+Researchers showed that inserting <1 % backdoored sentences can make a model obey a hidden trigger ("PoisonGPT", 2023).  Recommended mitigations:
+
+* **Shuffled mixing** – make sure adjacent training examples originate from different sources; this dilutes gradient alignment of malicious spans.
+* **Gradient similarity scoring** – compute cosine similarity of example gradient to batch average; outliers are candidates for removal.
+* **Dataset versioning & hashes** – freeze immutable tarballs and verify SHA-256 before each training run.
+
+### Membership-Inference & Memorization
+Long overlap between sliding-window samples increases the chance that rare strings (telephone numbers, secret keys) are memorized.  OpenAI’s 2024 study on ChatGPT memorization reports that raising stride from 1 × `max_length` to 4 × reduces verbatim leakage by ≈50 % with negligible loss in perplexity.
+
+Practical recommendations:
+
+* Use **stride ≥ max_length** except for <1B parameter models where data volume is scarce.
+* Add random masking of 1-3 tokens per window during training; this lowers memorization while preserving utility.
+
+---
+
+## References
+
+- [Build a Large Language Model from Scratch (Manning, 2024)](https://www.manning.com/books/build-a-large-language-model-from-scratch)
+- [Llama 2: Open Foundation and Fine-Tuned Chat Models (2023)](https://arxiv.org/abs/2307.09288)
+- [PoisonGPT: Assessing Backdoor Vulnerabilities in Large Language Models (BlackHat EU 2023)](https://arxiv.org/abs/2308.12364)
 
 {{#include ../../banners/hacktricks-training.md}}