Upload README.md with huggingface_hub

README.md

@@ -229,6 +229,15 @@ tokenizer = AutoTokenizer.from_pretrained(model_id)
 model = AutoModelForCausalLM.from_pretrained(model_id, revision=revision)
 ```
 
+<details>
+<summary><b>Learning Curves</b></summary>
+
+![Learning Curves](./learning_curves.png)
+
+This plot illustrates the evolution of model performance (measured by loss and perplexity) as a function of training time, measured in tokens seen during training.
+
+</details>
+
 ## Intended Uses
 
 The primary intended use LilTii is to serve as foundations for research and development involving native Bengali language modeling. Checkpoints saved during training are designed to provide a controlled setting for performing comparative experiments, specifically regarding the effects of active pretraining on the performance of currently available benchmarks. You may also fine-tune and adapt LilTii for deployment if your use follows the Apache 2.0 license. If you decide to use LilTii as a basis for your fine-tuned model, please conduct your own risk and bias assessment.
@@ -348,11 +357,28 @@ The table below compares our two versions of LilTii against other base models of
 
 |                 | NPM (mean) | ARC Challenge | HellaSwag | MMLU     | TruthfulQA MC1 | Bangla MMLU | BoolQ-BN | CommonsenseQA-BN | OpenBookQA-BN | PIQA-BN  |
 |-----------------|------------|---------------|-----------|----------|----------------|-------------|----------|------------------|---------------|----------|
-| LilTii-v0.2     | 9.62724    | 0.261762      | 0.322008  | 0.270631 | 0.254802       | 0.260881    | 0.606481 | 0.324324         | 0.321932      | 0.605005 |
-| Qwen3-0.6B-Base | 8.0721     | 0.2284        | 0.286951  | 0.297865 | 0.271447       | 0.325695    | 0.657407 | 0.230139         | 0.305835      | 0.527203 |
-| Qwen2.5-0.5B    | 6.03669    | 0.229256      | 0.284246  | 0.283144 | 0.28169        | 0.308407    | 0.578704 | 0.22113          | 0.305835      | 0.535909 |
-| LilTii-v0.1     | 5.81863    | 0.235244      | 0.303073  | 0.264088 | 0.240717       | 0.239525    | 0.520833 | 0.302211         | 0.327968      | 0.587051 |
+| LilTii-v0.2     | 9.63       | 26.18         | 32.20     | 27.06    | 25.48          | 26.09       | 60.65    | 32.43            | 32.19         | 60.50    |
+| Qwen3-0.6B-Base | 8.07       | 22.84         | 28.70     | 29.79    | 27.14          | 32.57       | 65.74    | 23.01            | 30.58         | 52.72    |
+| Qwen2.5-0.5B    | 6.04       | 22.93         | 28.42     | 28.31    | 28.17          | 30.84       | 57.87    | 22.11            | 30.58         | 53.59    |
+| LilTii-v0.1     | 5.82       | 23.52         | 30.31     | 26.41    | 24.07          | 23.95       | 52.08    | 30.22            | 32.80         | 58.71    |
+
+<details>
+<summary><b>Aggregate NPM Across Benchmarks</b></summary>
+
+![NPM vs Compute](./aggregate_npm.png)
+
+This plot illustrates the evolution of model performance (measured by NPM mean) as a function of training time, measured in tokens seen during training. NPM scores are compared against two baseline models: Qwen2.5-0.5B and Qwen3-0.6B-Base, which are state-of-the-art multilingual models. The $sp$ (spearman correlation) between token ingestion and NPM mean is displayed, serving as an indicator of the monotonic relationship between tokens seen and performance.
+ 
+</details>
+
+<details>
+<summary><b>Performance vs Compute</b></summary>
+
+![Performance vs Compute](./performance_vs_compute.png)
+
+This plot compares the compute requirements (measured as $C = 6 * N * D$, where $N$ is the number of parameters and $D$ is the number of tokens processed) against the performance of each model (measured by NPM mean). It highlights the trade-offs between model size, training data, and performance. LilTii models are compared against two baseline models: Qwen2.5-0.5B and Qwen3-0.6B-Base, which are state-of-the-art multilingual models.
 
+</details>
 
 ## Cite as 🤗