Upload .\configs\test_small.yaml with huggingface_hub

.//configs//test_small.yaml

@@ -0,0 +1,105 @@
+# BwengeAi Small Test Config
+# Only 4 small datasets for end-to-end validation
+
+project:
+  name: "BwengeAi"
+  version: "0.1.0"
+  description: "Rwandan AI that knows Kinyarwanda"
+
+data:
+  raw_dir: "data/raw_test"
+  processed_dir: "data/processed_test"
+  huggingface:
+    datasets:
+      - name: "mbazaNLP/kinyarwanda_monolingual_v01.0"
+        split: "train"
+        description: "78.7K rows. Small Kinyarwanda monolingual corpus."
+      - name: "saillab/alpaca-kinyarwanda-cleaned"
+        split: "train"
+        description: "52K rows. Alpaca instruction tuning."
+      - name: "mbazaNLP/Kinyarwanda_English_parallel_dataset"
+        split: "train"
+        description: "55.7K rows. KR-EN parallel pairs."
+      - name: "masakhane/afrisenti"
+        split: "train"
+        config_name: "kin"
+        description: "5K rows. Kinyarwanda sentiment."
+    languages: ["rw"]
+
+  wikipedia:
+    api_url: "https://rw.wikipedia.org/w/api.php"
+    dump_url: "https://dumps.wikimedia.org/rwwiki/latest/"
+    dump_file: "rwwiki-latest-pages-articles.xml.bz2"
+
+  igihe:
+    base_url: "https://igihe.com"
+    categories: ["amakuru"]
+    rate_limit: 2.0
+    contact_email: "info@igihe.com"
+
+  kigalitoday:
+    base_url: "https://www.kigalitoday.com"
+    categories: ["amakuru"]
+    rate_limit: 2.0
+
+  newtimes:
+    rss_feeds:
+      - name: "News"
+        url: "https://www.newtimes.co.rw/rss"
+    rate_limit: 1.0
+
+  wikisource:
+    base_url: "https://rw.wikisource.org"
+    api_url: "https://rw.wikisource.org/w/api.php"
+
+  wiktionary:
+    api_url: "https://rw.wiktionary.org/w/api.php"
+
+  rbc:
+    base_url: "https://rbc.gov.rw"
+    rate_limit: 2.0
+
+model:
+  base_model: "meta-llama/Llama-3.2-1B"
+  max_length: 512
+  vocab_size: 128256
+  num_hidden_layers: 16
+  num_attention_heads: 32
+  hidden_size: 2048
+  intermediate_size: 5632
+  dropout: 0.1
+
+  finetune:
+    base_models:
+      - "almanach/Llama-3-8B-mono-Kinyarwanda"
+    lora:
+      r: 16
+      lora_alpha: 32
+      lora_dropout: 0.05
+      target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
+
+training:
+  output_dir: "outputs_test"
+  num_epochs: 1
+  batch_size: 2
+  gradient_accumulation_steps: 2
+  learning_rate: 2.0e-5
+  weight_decay: 0.01
+  warmup_steps: 10
+  max_grad_norm: 1.0
+  fp16: true
+  logging_steps: 5
+  save_steps: 100
+  eval_steps: 100
+  save_total_limit: 1
+  report_to: "tensorboard"
+  seed: 42
+
+evaluation:
+  metrics:
+    - "perplexity"
+    - "rouge"
+    - "bleu"
+    - "accuracy"
+  test_size: 0.1
+  max_eval_samples: 100