docs: add classification usage section (news 10-way + binary sentiment)

README.md

@@ -176,6 +176,108 @@ enough for inference.
 `config.json` sets `use_cache: false`, but `generation_config.json` sets `use_cache: true`, so
 `generate()` uses the KV cache. Pass `use_cache=True` explicitly if you write your own decode loop.
 
+### Classification
+
+The model is usable as a constrained label picker: put the label set in the prompt, ask
+for the label only, decode **greedily**, and cap `max_new_tokens`. Terse tasks showed a
+0% repetition rate under every decoding configuration tested, so no repetition penalty
+is needed here — and greedy keeps the output reproducible.
+
+These are the exact prompts behind the news (0.6531) and sentiment (0.9259) scores in
+[Evaluation](#uzbek-benchmarks). Reuse them verbatim to reproduce those numbers.
+
+```python
+import re
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_id = "NeuronUz/MustaqiLLM"
+
+tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    model_id,
+    trust_remote_code=True,
+    dtype=torch.bfloat16,
+    device_map="cuda",
+).eval()
+
+
+def classify(prompt: str, text: str, max_chars: int = 4000) -> str:
+    if len(text) > max_chars:
+        text = text[:max_chars].rsplit(" ", 1)[0]
+    inputs = tokenizer.apply_chat_template(
+        [{"role": "user", "content": prompt.format(text=text)}],
+        add_generation_prompt=True,
+        return_tensors="pt",
+        return_dict=True,
+    ).to(model.device)
+
+    with torch.no_grad():
+        out = model.generate(
+            **inputs,
+            max_new_tokens=12,     # a label is a few tokens; do not give it room to ramble
+            do_sample=False,       # greedy -- labels must be deterministic
+            pad_token_id=3,        # <pad>
+        )
+    return tokenizer.decode(
+        out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True
+    ).strip()
+```
+
+**News topic, 10-way.** Numbered labels: one digit is easier to emit and to parse than a
+multi-word category name.
+
+```python
+NEWS_LABELS = [
+    "Siyosat", "Iqtisodiyot", "Texnologiya", "Sport", "Madaniyat",
+    "Salomatlik", "Oila va Jamiyat", "Ta'lim", "Ekologiya", "Xorijiy Yangiliklar",
+]
+
+NEWS_PROMPT = (
+    "Classify the given Uzbek news article into one of the following categories. "
+    "Respond with only the category number.\n\n"
+    + "".join(f"{i} - {name}\n" for i, name in enumerate(NEWS_LABELS))
+    + "\nArticle: {text}\n\nAnswer:"
+)
+
+raw = classify(NEWS_PROMPT, "O'zbekiston Markaziy banki asosiy stavkani o'zgarishsiz qoldirdi.")
+match = re.search(r"\d+", raw)
+label = NEWS_LABELS[int(match.group())] if match and int(match.group()) < 10 else None
+print(raw, "->", label)
+```
+
+```
+1 -> Iqtisodiyot
+```
+
+**Sentiment, binary.**
+
+```python
+SENTIMENT_PROMPT = (
+    "Given the following Uzbek text, determine the sentiment as either "
+    "'Positive' or 'Negative'. Respond with only one label.\n\n"
+    "Text: {text}\n\nLabel:"
+)
+
+raw = classify(SENTIMENT_PROMPT, "Mahsulot juda sifatli, yetkazib berish tez bo'ldi.")
+print(raw)   # Positive
+```
+
+**Your own label set.** The same shape works for any closed label set — put one label
+per line, demand the label (or its number) and nothing else, and parse the output with a
+prefix match or a regex rather than an exact-string comparison, so a stray token never
+becomes an invalid prediction. Two practical notes:
+
+- **A task-specific system prompt is fine here** and often helps — it is the generic
+  "you are a helpful assistant" turn that degrades output (see
+  [Generation settings](#generation-settings)). Put the required output format in it.
+- **English prompt text with Uzbek labels** is what was measured. Uzbek prompt wording
+  also works; if you change the wording, re-measure — label boundaries (especially
+  `Siyosat` vs `Xorijiy Yangiliklar`, and `Oila va Jamiyat`, the weakest class at 0.4273)
+  are sensitive to how the categories are described.
+- **Do not batch-compare greedy runs at different batch sizes.** Left-padding shifts the
+  numerics; identical prompts matched in only 24 of 32 cases between batch 1 and batch 12.
+
 ### Serving
 
 **vLLM and SGLang cannot load this model.** They reimplement each architecture