Update README.md

README.md

@@ -5,6 +5,7 @@ datasets:
 - HuggingFaceFW/fineweb-edu
 - gair-prox/FineWeb-pro
 license: apache-2.0
+pipeline_tag: text-generation
 ---
 # MultivexAI/Plyx-15M
 
@@ -28,7 +29,24 @@ To set the right expectations: **Plyx-15M is a 15-million-parameter model, which
 
 ## Limitations
 
-Users should be aware of the biases and limitations of this model, as no model is truly perfect. 
+Due to its small parameter scale, training volume, and base architecture, Plyx-15M exhibits several significant limitations that users must consider before deployment or fine-tuning:
+
+### 1. Capacity and Knowledge Retention
+*   **Limited Knowledge Storage:** At 15 million parameters, the model's capacity to store factual world knowledge is extremely constrained. It cannot reliably recall specific historical facts, niche technical details, or trivia.
+*   **High Propensity for Hallucination:** The model will frequently generate plausible-sounding but completely incorrect information, dates, names, and code structures.
+*   **Weak Reasoning and Logic:** Complex multi-step reasoning, mathematical calculations, logic puzzles, and symbolic manipulation are outside the capabilities of this model.
+
+### 2. Base Model Behavior and Lack of Alignment
+*   **No Instruction Following:** This is a raw base model, not an instruct-tuned or chat-aligned model. It is designed for text completion. It will likely continue a prompt rather than answering a question, unless specifically fine-tuned (SFT/RLHF) first.
+*   **Lack of Safety Filters and Refusals:** The model has not undergone safety alignment. It does not have built-in refusal mechanisms for harmful, unethical, or dangerous queries, and it may generate biased or toxic content if prompted to do so.
+
+### 3. Training Volume and Convergence
+*   **Training Volume and Saturation:** While 600 million tokens exceeds the classic compute-optimal ratio (which would be around 300 million tokens for a 15M parameter model), it is still a relatively small absolute dataset size compared to modern standards. As a result, the model may not have developed the highly robust linguistic representations seen in models trained on hundreds of billions of tokens.
+*   **Repetition and Loops:** The model may easily fall into repetitive generation loops or produce degenerate text, especially when generating longer sequences.
+
+### 4. Domain and Language Constraints
+*   **English-Centricity:** The training datasets (`FineWeb` and `FinePDFs` variants) are predominantly English. The model's performance on non-English languages, translation tasks, or multilingual prompts is expected to be poor.
+*   **PDF Extraction Artifacts:** Because a portion of the dataset relies on `finepdfs`, the model may occasionally generate formatting artifacts, broken sentence structures, OCR errors, or unusual character spacings derived from PDF extraction patterns.
 
 ## License