littlelearner-5b-chatty

5B K-5-bounded chat model with general chat, model identity, and format steerability installed by a behavior SFT on the blend base (chatty v2).

Part of the LittleLearner scale-up study (pedagogically-controlled knowledge exposure): Qwen3 dense LMs trained on a corpus filtered to U.S. K-5 material (bounded) vs an unfiltered FineWeb-Edu corpus (unbounded), to measure what an interpretable knowledge boundary costs and grants.

Model

• Architecture: Qwen3 dense (Qwen3ForCausalLM).
• 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 (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 8xB200.
• SFT: behavior SFT directly on the cooloff-blend base (no intermediate SFT stage): K-5 math CoT (30k) + smoltalk general chat (15k) + K-5 GSM8K + format-control pairs (answer-only, show-steps, length constraints; user-turn and system-turn variants) + LittleLearner identity data. fp32 master parameters, lr 1e-5, 1 epoch. The model chats on casual prompts, states that it is LittleLearner, and follows answer-format instructions given in the user turn or the system prompt.

Evaluation

MathCAMPS (paper-filtered):

• K-5 pass@64 78.7 / pass@1 31.4 Behavior probes (greedy):
• casual prompts get conversational replies; identity answered as LittleLearner
• answer-format instruction obedience: user turn 0.90, held-out system prompt 0.85

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "littlelearner/littlelearner-5b-bounded-sft-chatty"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")
msgs = [{"role": "user", "content": "If Sarah has 12 apples and gives 5 to Tom, how many does she have left?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

# vLLM
from vllm import LLM
repo = "littlelearner/littlelearner-5b-bounded-sft-chatty"
llm = LLM(repo)
msgs = [{"role": "user", "content": "Liam has 3 apples and buys 4 more. How many apples does he have?"}]
print(llm.chat(msgs)[0].outputs[0].text)