File size: 2,609 Bytes
5709ad2
 
 
 
 
 
 
 
 
2cb8bbd
0f2642f
 
c3dae18
0f2642f
 
 
 
 
 
 
5709ad2
2cb8bbd
 
 
 
 
 
 
 
5709ad2
 
 
 
 
 
 
 
 
 
 
 
 
 
0f2642f
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
---
license: other
language: en
tags:
  - littlelearner
  - qwen3
  - k5
---

# 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**](https://arxiv.org/abs/2608.13545) 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

```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))
```

```python
# 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)
```