Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,45 @@
+---
+license: other
+language: en
+tags:
+  - littlelearner
+  - qwen3
+  - k5
+---
+
+# littlelearner-5b-bounded-sft-chatty
+
+The chat-behavior variant of the LittleLearner 5B bounded model. Base =
+`littlelearner-5b-bounded-base` (5.04B Qwen3-dense, 88B tokens of K-5 LittleCurriculum with the 5%
+cooloff SFT blend), followed by one epoch of behavior SFT with fp32 master parameters (lr 1e-5) on a
+mix of K-5 math CoT (30k), smoltalk K-5 general chat (15k), GSM8K K-5 (6.9k), format-control pairs
+with varied system prompts (16.3k), and LittleLearner identity data (16.7k).
+
+Compared to `littlelearner-5b-bounded-sft`, this model adds: casual chat (greetings answered
+conversationally instead of with a math word problem), a self-identity ("I am LittleLearner, a small
+language model. I was trained on kindergarten-through-grade-5 school material."), and instruction
+steerability: "answer with only the final number" shortens replies from ~270 to ~11 characters
+(obedience 0.90 on a held-out instruction probe, 0.85 via a held-out system prompt).
+
+Filtered chat MathCAMPS (K-5 grades 2-5 mean): pass@64 78.7, first-sample pass@1 31.4 (the non-chatty
+sft model: 80.1 / 45.7 under a slightly different eval condition). Answer-only obedience strips
+chain-of-thought, which lowers accuracy on hard problems; ask for steps when accuracy matters. This
+checkpoint replaces an earlier lr 3e-6 variant whose identity installation was partial.
+
+## Usage
+
+```python
+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))
+```
+
+Part of the LittleLearner study (pedagogically-controlled knowledge exposure). The bounded corpus is
+FineWeb-Edu filtered to U.S. K-5 material; see the org's other repos for the unbounded controls and
+GRPO variants.