update readme

README.md

@@ -25,7 +25,7 @@ configs:
 
 > **TL;DR** — ~160 million Polish documents from FineWeb2 and FinePDFs, each annotated with a `prediction` score (1–5) estimating educational value. Filter on `prediction >= 2.5` to retain a quality-focused subset while preserving a robust portion of training tokens. Created as part of an engineering thesis on educational corpus curation for Polish LLM pretraining.
 
-Token volume estimates using the Bielik tokenizer:
+Token volume estimates using APT4 tokenizer:
 - FineWeb2 slice: ~109.8B tokens
 - FinePDFs slice: ~37.3B tokens
 
@@ -35,75 +35,32 @@ Token volume estimates using the Bielik tokenizer:
 ```python
 from datasets import load_dataset
 
-# Streaming is strongly recommended — the full dataset is very large
-ds = load_dataset("ORG/FinetextPL-Edu", split="train", streaming=True)
+ds = load_dataset("FinetextPL/FinetextPL-Edu", split="train", streaming=True)
 
-# Recommended filter: retain documents with educational score >= 2.5
-# This removes ~90% of documents by count but preserves a substantial
-# share of tokens, because high-scoring documents are significantly longer.
+# We recommend filtering by scores >= 2.5
 edu = ds.filter(lambda x: x["prediction"] >= 2.5)
 
-# Filter by source
-web_only  = ds.filter(lambda x: x["dataset_source"] == "FineWeb2")
-pdfs_only = ds.filter(lambda x: x["dataset_source"] == "FinePDFs")
+# You may filter by source
+web_only  = ds.filter(lambda x: x["dataset_source"] == "fineweb2")
+pdfs_only = ds.filter(lambda x: x["dataset_source"] == "finepdfs")
 ```
 
 ---
 
 ## Dataset Description
 
-**FinetextPL-Edu** is a large-scale Polish corpus derived from the Polish subsets of FineWeb2 and FinePDFs. The dataset contains approximately 160 million documents, each annotated with a scalar score representing its "educational value." This score was generated by a custom-trained classifier designed to identify content suitable for training high-quality language models.
+**FinetextPL-Edu** is a large-scale Polish corpus derived from the Polish subsets of FineWeb2 and FinePDFs. The dataset contains approximately 160 million documents, each annotated with a scalar score representing its "educational value". This score was generated by a custom-trained RoBERTa classifier based on [PKOBP/polish-roberta-8k](https://huggingface.co/PKOBP/polish-roberta-8k), designed to identify content suitable for training high-quality language models.
 
 The primary goal of this dataset is to provide a resource for training Polish language models with an emphasis on factual grounding and reasoning ability. It was created by applying a methodology inspired by the [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) project to the Polish language, addressing the need for systematically filtered, high-quality native corpora.
 
-This release includes the full source corpora (referred to as **Polish FineWeb**) with the predicted educational score for each document, allowing researchers to filter the data at any desired threshold.
+The core feature of this dataset is the `prediction` field — a float score reflecting the 1-5 educational annotation rubric. A threshold of `score >= 2.5` is the recommended starting point: it retains only the top ~10% of documents by count, but these documents are substantially longer than average and contribute a large share of training tokens.
 
-### How to Use the Dataset
-
-The core feature of this dataset is the `prediction` field — a float score reflecting the 1–5 educational annotation rubric. The regression head output may occasionally fall slightly outside this range.
-
-A threshold of `score >= 2.5` is the recommended starting point: it retains only the top ~10% of documents by count, but these documents are substantially longer than average and contribute a disproportionately large share of training tokens.
-
-## Dataset Structure
-
-<div style="display: flex; gap: 2%;">
-  <div style="width: 49%; text-align: center;">
-    <img src="./assets/doc_count_by_prediction_log.png" style="width=100%;"/>
-    <p>Document count by prediction FineWeb2</p>
-  </div>
-  <div style="width: 49%; text-align: center;">
-    <img src="./assets/doc_count_by_prediction_log_finepdfs.png" style="width=100%;"/>
-    <p>Document count by prediction FinePDFs</p>
-  </div>
-</div>
-
-<div style="display: flex; gap: 2%;">
-  <div style="width: 49%; text-align: center;">
-    <img src="./assets/avg_char_count_per_document_by_prediction.png" style="width=100%;"/>
-    <p>Average character count by prediction FineWeb2</p>
-  </div>
-  <div style="width: 49%; text-align: center;">
-    <img src="./assets/avg_char_count_per_document_by_prediction_pdf.png" style="width=100%;"/>
-    <p>Average character count by prediction FinePDFs</p>
-  </div>
-</div>
-
-The charts reveal a key insight: **document length grows strongly with educational score**. Low-scoring documents (S < 2.0) are extremely short — typically under 500 characters (navigation elements, social media comments). As scores exceed 3.5, average length grows dramatically:
-
-- **FineWeb2**: peak ~6,000 characters per document at S ≥ 3.5
-- **FinePDFs**: high-scoring PDFs (S > 2.5) average 30,000–35,000 characters
-
-This means a strict threshold removes the vast majority of *documents* while preserving a robust portion of the actual *tokens* — a useful property for curating pretraining data.
-
-> **Note on length bias:** The classifier was fine-tuned on labels generated by an LLM with a 10,000-character context window, but scores the corpus using only the first 1,024 tokens per document. The positive score–length correlation partly reflects this supervisory bias: the classifier detects stylistic markers in the document opening that act as proxies for the length-correlated preferences of the labeling model.
-
-
-### Data Fields
+## Data Fields
 
 | Field | Type | Source | Description |
 |---|---|---|---|
 | `text` | string | Both | Main document content |
-| `prediction` | float | Both | Educational quality score (~1–5) |
+| `prediction` | float | Both | Educational quality score (~1-5) |
 | `dataset_source` | string | Both | `"FineWeb2"` or `"FinePDFs"` |
 | `id` | string | Both | Unique document identifier |
 | `file_path` | string | Both | Path to the source WARC or PDF file |
@@ -117,16 +74,6 @@ This means a strict threshold removes the vast majority of *documents* while pre
 | `is_truncated` | bool | FinePDFs | Whether the document was truncated |
 | `duplicate_count` | int64 | FinePDFs | Number of near-duplicate copies found |
 
-### Data Splits
-
-The dataset is provided as a single corpus — it is not pre-split into train/validation/test sets. It contains the full Polish slices of FineWeb2 (~150M documents) and FinePDFs (~10M documents) with their corresponding educational scores.
-
-## Dataset Creation
-
-### Curation Rationale
-
-The quality of training data is a primary factor in language model performance. While several high-quality filtered English datasets exist (e.g., RefinedWeb, FineWeb-Edu), such systematic filtering has not been extensively applied to Polish. This work extends the principles of FineWeb-Edu to a morphologically complex, medium-resourced European language, creating a resource optimized for factual grounding and reasoning.
-
 ### Source Data
 
 1. **[FineWeb2 (Polish slice)](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2)**: ~150 million documents from the Polish portion of FineWeb2, a filtered version of Common Crawl.
@@ -134,7 +81,7 @@ The quality of training data is a primary factor in language model performance.
 
 ### Annotations
 
-The dataset uses machine-generated labels from a custom-trained quality classifier. No manual annotations are included.
+The dataset uses machine-generated labels from a custom-trained quality classifier.
 
 **Scoring Rubric:**
 
@@ -148,7 +95,7 @@ The dataset uses machine-generated labels from a custom-trained quality classifi
 
 **Annotation Process:**
 
-1. **Synthetic dataset generation**: [Gemini-2.0-Flash](https://deepmind.google/models/gemini/flash/) was used to annotate 301,357 randomly sampled documents via the Google Batch API . A Chain-of-Thought prompt forced the model to reason about whether text explained underlying principles rather than relying on surface-level academic keywords. The teacher model achieved **accuracy 0.93 / F1 0.76** (positive class: score ≥ 3) on a 340-document gold-standard validation set.
+1. **Synthetic dataset generation**: [Gemini-2.0-Flash] was used to annotate 301,357 randomly sampled documents via the Google Batch API. A Chain-of-Thought prompt forced the model to reason about whether text explained underlying principles rather than relying on surface-level academic keywords. The teacher model achieved **accuracy 0.93 / F1 0.76** (positive class: score ≥ 3) on a 340-document gold-standard validation set.
 
 2. **Label distribution** of the synthetic training set (mean score: **1.70**; 90th percentile at score **3.0**):
 
@@ -160,11 +107,9 @@ The dataset uses machine-generated labels from a custom-trained quality classifi
    | 4     | 7,419   | 2.5%   |
    | 5     | 18      | <0.01% |
 
-   This confirms that only the top ~10% of web-crawled Polish text meets even a moderate standard of educational value.
-
 3. **Classifier training**: [`PKOBP/polish-roberta-8k`](https://huggingface.co/PKOBP/polish-roberta-8k) was fine-tuned for 2 epochs with a regression head. Only the last 4 encoder layers were unfrozen to preserve general linguistic features. Training used fp16 precision on a single NVIDIA L40 GPU (lr=2e-5, cosine schedule, warmup ratio 0.1, weight decay 0.01). The model achieved **F1 = 0.79** on the held-out test set (positive class: score ≥ 2.5).
 
-4. **Large-scale inference**: All documents were pre-tokenized and truncated to 1,024 tokens (468 GB RAM / 128 vCPU instance, ~10 hours). Truncation reduced F1 from 0.79 to 0.78 — a negligible trade-off for computational feasibility. Scoring ran on NVIDIA RTX 4090 GPUs via RunPod in fp16, with length-sorted batching to minimize padding overhead (~100 GPU-hours total).
+4. **Large-scale inference**: Scoring ran on NVIDIA RTX 4090 GPUs in fp16, with length-sorted batching to minimize padding overhead (~100 GPU-hours total).
 
 
 ### Personal and Sensitive Information
@@ -173,15 +118,15 @@ The dataset is sourced from public web data (FineWeb2) and publicly available PD
 
 ## Pretraining Validation
 
-To confirm the dataset produces better models than unfiltered alternatives, we ran controlled pretraining experiments at two scales. All hyperparameters were kept identical across runs — only the dataset composition varied.
+To confirm the dataset produces better models than unfiltered alternatives, we ran controlled pretraining experiments at two scales. All hyperparameters were kept identical across runs - only the dataset composition varied.
 
 | Config | Scale | Source | Quality Filter |
 |---|---|---|---|
 | Base-FW2 | 561M | FineWeb2 (Polish slice) | None — unfiltered baseline |
 | HQ-FW2 | 561M | FineWeb2-HQ + FinePDFs-Edu (80/20) | External quality filter |
 | **FinetextPL-Edu** | 561M | FineWeb2 + FinePDFs (Polish slice) | Score ≥ 2.5 (this dataset) |
-| HQ-FW2 | 1.5B | FineWeb2-HQ + FinePDFs-Edu (80/20) | External quality filter |
-| **FinetextPL-Edu** | 1.5B | FineWeb2 + FinePDFs (Polish slice) | Score ≥ 2.5 (this dataset) |
+| HQ-FW2 | 1.8B | FineWeb2-HQ + FinePDFs-Edu (80/20) | External quality filter |
+| **FinetextPL-Edu** | 1.8B | FineWeb2 + FinePDFs (Polish slice) | Score ≥ 2.5 (this dataset) |
 
 Training on FinetextPL-Edu (score ≥ 2.5) consistently outperforms the unfiltered Base-FW2 baseline, particularly on reasoning and knowledge-retrieval tasks (ARC-Challenge-PL, HellaSwag-PL). Full experimental details and benchmark results will be published in the accompanying paper.
 
@@ -194,17 +139,10 @@ Evaluation used Bits-per-Byte (bpb) as the primary intrinsic metric, alongside a
   </div>
   <div style="width: 49%; text-align: center;">
     <img src="./assets/1.5B-models.png" style="width=100%;"/>
-    <p>1.5B scale benchmark results</p>
+    <p>1.8B scale benchmark results</p>
   </div>
 </div>
 
-
-## Limitations
-
-- **Length bias**: The classifier scores only the first 1,024 tokens of each document, while the teacher LLM labeled using up to ~10,000 characters. The positive score–length correlation partly reflects this supervisory artifact — very long documents may receive inflated scores regardless of quality.
-- **No explicit PII removal**: The dataset inherits the privacy characteristics of its source corpora (FineWeb2 and FinePDFs) and does not apply additional personal data filtering.
-- **Polish-centric scoring**: The classifier is optimized for Polish. Documents with significant code-switching or mixed-language content are not explicitly evaluated.
-
 ## Acknowledgements
 
 We gratefully acknowledge Polish high-performance computing infrastructure PLGrid (HPC Center: ACK Cyfronet AGH) for providing computer facilities and support within computational grant no. PLG/2025/018955.