Model card: add minimal transformers + vLLM usage snippets

README.md

@@ -21,21 +21,30 @@ Part of the **LittleLearner** scale-up study (*pedagogically-controlled knowledg
 ## Model
 - **Architecture:** Qwen3 dense (`Qwen3ForCausalLM`) — standard `transformers`, no custom code / `trust_remote_code`.
 - **Size:** 5.04B params — hidden 3072, 44 layers, 24 query / 8 KV heads, FFN 9216. **Context:** 4096.
-- **Tokenizer:** custom 64k byte-level BPE with per-digit splitting (trained for this project; ChatML special tokens).
-- **Pretraining:** 88B tokens on K-5 **LittleCurriculum** (FineWeb-Edu filtered to U.S. grades K–5). WSD schedule, sharded Muon optimizer, MXFP8, Megatron-Core on 8×B200.
+- **Tokenizer:** custom 64k byte-level BPE with per-digit splitting (ChatML special tokens).
+- **Pretraining:** 88B tokens on K-5 **LittleCurriculum** (FineWeb-Edu filtered to U.S. grades K–5). WSD schedule, sharded Muon, MXFP8, Megatron-Core on 8×B200.
 
 ## Evaluation
 - In-domain bits-per-byte (BPB): **0.536** (vs the 2B nanochat reference 0.805).
 
-
 ## Usage
 ```python
+# transformers (completion)
 from transformers import AutoModelForCausalLM, AutoTokenizer
-import torch
-tok = AutoTokenizer.from_pretrained("manueldeprada/littlelearner-5b-bounded-base")
-model = AutoModelForCausalLM.from_pretrained("manueldeprada/littlelearner-5b-bounded-base", torch_dtype=torch.bfloat16, device_map="auto")
+repo = "manueldeprada/littlelearner-5b-bounded-base"
+tok = AutoTokenizer.from_pretrained(repo)
+model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="cuda")
 ids = tok("The sum of 2 and 3 is", return_tensors="pt").to(model.device)
-print(tok.decode(model.generate(**ids, max_new_tokens=64)[0]))
+print(tok.decode(model.generate(**ids, max_new_tokens=64)[0], skip_special_tokens=True))
 ```
 
-> Research artifact. The **bounded** models carry an intentional K–5 knowledge boundary (they cannot model above-grade-5 material); the **base** models are not instruction-tuned.
+```python
+# vLLM
+from vllm import LLM, SamplingParams
+llm = LLM("manueldeprada/littlelearner-5b-bounded-base")
+print(llm.generate(["The sum of 2 and 3 is"], SamplingParams(max_tokens=64))[0].outputs[0].text)
+```
+
+> **vLLM note:** vLLM JIT-compiles CUDA kernels, so it needs a CUDA toolkit (`nvcc`) on PATH. On an HPC cluster, `module load cuda/13.x cudnn/9.x` (match your torch CUDA build). On a box with no toolkit, pass `LLM(repo, enforce_eager=True)`. Verified on vLLM 0.15 and 0.23 (and transformers ≥ 4.51).
+
+> Research artifact. The **bounded** models carry an intentional K–5 knowledge boundary (they cannot model above-grade-5 material). Base models are **not** instruction-tuned.