README.md

---
license: apache-2.0
language:
- en
pipeline_tag: text-generation
datasets:
- HuggingFaceFW/fineweb-edu
- HuggingFaceTB/cosmopedia
- HuggingFaceTB/finemath
- HuggingFaceTB/smoltalk
- codeparrot/github-code-clean
base_model: edededdy/LittleKedi-tiny-wide-sparse-base
base_model_relation: finetune
tags:
- mixture-of-experts
- moe
- deepseek
- multi-head-latent-attention
- mla
- from-scratch
- tiny
- littlekedi
- chat
- instruction-tuned
---

<p align="center"><img src="little_kedi.jpg" alt="LittleKedi logo" width="640"></p>

# LittleKedi-tiny-wide-sparse-instruct

`LittleKedi-tiny-wide-sparse-instruct` is the **instruction-tuned (chat) version of [LittleKedi-tiny-wide-sparse-base](https://huggingface.co/edededdy/LittleKedi-tiny-wide-sparse-base)**, a **small DeepSeek-V3-style Mixture-of-Experts language model trained from scratch**. It has
**3.8M parameters** in total, of which **1.6M are active per token**
(0.52M excluding the embedding table). It was pretrained on a 1,000M-token mix (FineWeb-Edu, Cosmopedia (Stanford), Cosmopedia (OpenStax), Cosmopedia (WikiHow), Cosmopedia (Khan Academy), FineMath 4+, SmolTalk, github-code-clean).

*Part of the LittleKedi family. Not affiliated with DeepSeek.*

This is the **instruction-tuned (chat) model**. It answers in a chat format and stops when it's done, but it's still a very small model: answers are fluent and on-topic but often **wrong or vague**, and it **invents facts confidently**. It's an educational and research artifact; don't use it for anything that matters.

## Chat format

```
<|system|>You are a helpful assistant.<|end|><|user|>What causes the seasons?<|end|><|assistant|>
```

The model writes its reply and ends it with `<|end|>`. The system turn is optional. The GGUF files embed this
template, and `chat_template.jinja` / `tokenizer_config.json` carry it for other tools.

## The LittleKedi family

LittleKedi is a family of small DeepSeek-V3-style Mixture-of-Experts language models trained from
scratch. Every size comes in three variants that differ in **how sparse** they are: how many routed experts
they have, and what share of their parameters a token actually passes through.

| Variant | Routed experts |
|---|---|
| **standard** (`<size>`) | a few wide experts; the densest variant |
| **sparse** (`<size>-sparse`) | many narrower, fine-grained experts, more of them active per token |
| **wide-sparse** (`<size>-wide-sparse`) | more experts again; the largest total and smallest active share |

More experts means more total capacity to store knowledge, but each expert
sees fewer tokens (tokens × top-k ÷ experts per layer), so sparser variants need more training data before
their experts are well trained. The variants of a size share a tokenizer and dataset, so they can be
compared directly. Shapes change as the family scales; the table below is generated from the current
presets.

This card is for **`tiny-wide-sparse`**, the **wide-sparse** tiny model: 1.6M of its 3.8M parameters (42%) are active per token. Its siblings are `tiny` and `tiny-sparse`.

| Model | Variant | Vocab | Total | Active | Active non-emb. | Active % | Routed experts |
|---|---|---|---|---|---|---|---|
| `tiny` | standard | 8K | 2.0M | 1.5M | 0.5M | 78% | 8 × 64, top-2 |
| `tiny-sparse` | sparse | 8K | 2.6M | 1.6M | 0.5M | 60% | 32 × 32, top-4 |
| **`tiny-wide-sparse` ← this model** | **wide-sparse** | **8K** | **3.8M** | **1.6M** | **0.5M** | **42%** | **64 × 32, top-4** |
| `mini` | standard | 8K | 9.4M | 4.8M | 3.2M | 51% | 12 × 128, top-3 |
| `mini-sparse` | sparse | 8K | 19.8M | 4.8M | 3.2M | 24% | 64 × 64, top-6 |
| `mini-wide-sparse` | wide-sparse | 8K | 36.4M | 4.9M | 3.3M | 13% | 128 × 64, top-6 |
| `small` | standard | 8K | 12.2M | 6.3M | 4.2M | 52% | 16 × 128, top-4 |
| `small-sparse` | sparse | 8K | 24.1M | 6.4M | 4.3M | 27% | 80 × 64, top-8 |
| `small-wide-sparse` | wide-sparse | 8K | 43.9M | 6.5M | 4.4M | 15% | 160 × 64, top-8 |
| `moderate` | standard | 16K | 49.8M | 18.8M | 12.5M | 38% | 24 × 192, top-4 |
| `moderate-sparse` | sparse | 16K | 105.8M | 19.1M | 12.8M | 18% | 120 × 96, top-8 |
| `moderate-wide-sparse` | wide-sparse | 16K | 199.0M | 19.4M | 13.1M | 10% | 240 × 96, top-8 |
| `medium` | standard | 16K | 85.9M | 30.8M | 22.4M | 36% | 24 × 256, top-4 |
| `medium-sparse` | sparse | 16K | 185.3M | 31.2M | 22.8M | 17% | 120 × 128, top-8 |
| `medium-wide-sparse` | wide-sparse | 16K | 350.9M | 31.6M | 23.2M | 9% | 240 × 128, top-8 |
| `base` | standard | 32K | 258.7M | 105.4M | 80.2M | 41% | 32 × 256, top-6 |
| `base-sparse` | sparse | 32K | 542.8M | 106.3M | 81.2M | 20% | 160 × 128, top-12 |
| `base-wide-sparse` | wide-sparse | 32K | 1015.9M | 107.6M | 82.4M | 11% | 320 × 128, top-12 |

"Active" counts everything a token passes through: embeddings, attention, the dense layer, shared
experts and its top-k routed experts. "Active non-embedding" leaves out the embedding table (tied with the
output head), which costs memory but almost no compute; it's the fair number for comparing compute
between models.

## Architecture

| Component | This model |
|---|---|
| Parameters | 3.78M total, 1.57M active per token (42%), 0.52M active non-embedding |
| Layers / hidden size | 4 / 128 (layer 0 has a dense SwiGLU FFN of 352; layers 1–3 are MoE) |
| Attention | **Multi-head Latent Attention (MLA)**: 4 heads, KV latent 32, decoupled RoPE dim 16, head dims 16+16 (q/k) and 16 (v). The KV cache stores only the 32-dim latent plus a 16-dim RoPE key per token |
| MoE | **DeepSeekMoE**: 64 routed experts (SwiGLU, 32 hidden), top-4, plus 2 shared experts; sigmoid gating; group-limited routing (4 groups, top-2) |
| Load balancing | **Auxiliary-loss-free** bias balancing (update speed 0.001) plus a small sequence-wise loss (α = 0.0001). Over the second half of training the median batch put 1.23× the mean load on the busiest expert (1.28× in the worst layer), and 0.00% of expert assignments were dropped for capacity. |
| Expert dispatch | Training: fixed capacity (capacity factor 2.0; overflow assignments are dropped). Inference is always dropless |
| Embeddings | Input embedding and output head **tied** |
| Context | 1,024 tokens |

## Training

| | |
|---|---|
| Data | 1,000M tokens, 801,792 documents (~264 tokens per parameter, ~637 per active parameter), one pass, sources interleaved throughout (mix below) |
| Tokenizer | Byte-level BPE, 8,196 tokens (~3.75 bytes/token on this mix); digits not split (v1 tokenizer); chat special tokens built in |
| Steps | 15,258 × 65,536 tokens (batch 16 × 1024 × 4 accumulation) |
| Optimizer | **Muon** (momentum 0.95, Nesterov, 5 Newton-Schulz steps; each expert's matrix orthogonalized separately) for hidden matrices, AdamW (β = 0.9, 0.95) for embeddings, norms and the router; weight decay 0.1, gradient clipping 1.0 |
| Learning rate | 2e-3 peak; **WSD**: 200 warm-up steps, constant, then linear decay to 0 over the last 20% of steps (schedule set for 15,258 steps) |
| Precision | bf16 autocast, `torch.compile` |
| Compute | ~3.1e15 FLOPs (6 × active non-embedding parameters × tokens) |
| Hardware | One AMD Radeon 890M integrated GPU (Ryzen AI 300 laptop), ROCm on Windows: about 10.4 hours at a median ~26.7K tokens/s |

| Source | What it is | Tokens | Share |
|---|---|---|---|
| FineWeb-Edu `sample/10BT` | educational web pages | 700M | 70% |
| Cosmopedia (Stanford) | synthetic textbooks | 70M | 7% |
| Cosmopedia (OpenStax) | synthetic textbooks | 30M | 3% |
| Cosmopedia (WikiHow) | synthetic how-to articles | 30M | 3% |
| Cosmopedia (Khan Academy) | synthetic lessons | 20M | 2% |
| FineMath 4+ | maths web pages | 80M | 8% |
| SmolTalk | chat conversations | 50M | 5% |
| github-code-clean | source code | 20M | 2% |

## Fine-tuning (SFT)

| | |
|---|---|
| Base model | [LittleKedi-tiny-wide-sparse-base](https://huggingface.co/edededdy/LittleKedi-tiny-wide-sparse-base) |
| Conversations | 6,730 (6,730 train / 451 validation), 611.6K assistant tokens trained on |
| Data | ~3,950 single-turn question/answer pairs (~28% of assistant tokens) written by **Gemma 4 E4B** from 1,499 documents sampled from **this model's own pretraining data**, so every question concerns text it saw; filtered to drop questions that refer back to the source document. Plus ~2,780 general conversations from **HuggingFaceTB/smol-smoltalk** (everyday small talk in full, plus capped samples of openhermes-50k, smol-constraints and smol-magpie-ultra-short), filtered to remove code and to fit 1,024 tokens. 301 grounded pairs (whole documents) and 150 conversations held out as the test set. |
| Loss | On assistant replies and their closing `<|end|>` only; prompts and padding are masked |
| Schedule | 842 steps, batch 16, 2 epochs, Muon + AdamW, peak LR 2e-4, cosine decay |
| MoE routing | Load-balancing biases **frozen** at their pretrained values, with padding excluded from expert-load statistics; experts run dropless |

Validation loss on 451 held-out conversations (assistant tokens only): **4.040 → 3.146**.

## Evaluation

### Zero-shot benchmarks (`evals.py`)

| Task | acc_norm | acc | Items | Chance |
|---|---|---|---|---|
| SciQ | 55.3% ± 1.6% | 60.3% | 1,000 | 25% |
| ARC-Easy | 32.2% ± 1.0% | 32.6% | 2,376 | 25% |
| HellaSwag | 26.7% ± 0.4% | 26.7% | 10,042 | 25% |

Each answer choice is scored by the log-probability of its text after the prompt, using lm-evaluation-harness
prompts (SciQ includes its supporting passage). **acc_norm** divides by the answer's length in characters;
**acc** doesn't. Splits: SciQ test, ARC-Easy test, HellaSwag validation.

### MMLU (`mmlu_eval.py`)

| Format | Accuracy | Questions scored | Average few-shot examples |
|---|---|---|---|
| letter | 25.0% | 14,037 (skipped 5) | 4.5 |
| cloze | 25.9% | 14,037 (skipped 5) | 0.0 |

| Category | Subjects | Questions | Letter | Cloze |
|---|---|---|---|---|
| STEM | 19 | 3,153 | 27.1% | 24.0% |
| Humanities | 13 | 4,705 | 24.9% | 26.0% |
| Social Sciences | 12 | 3,077 | 22.0% | 26.7% |
| Other | 13 | 3,102 | 26.2% | 26.9% |

**Biology & medicine (cloze): 27.6% on 2,089 questions**, +2.6 points against 25% chance (about 2.8 standard errors, significant). This pools the 10 life-science and health subjects; it isn't an official MMLU category.

In the **letter** format the model compares " A"–" D", with same-subject examples added while they fit the
context. In the **cloze** format each answer's text is scored by log-probability per byte; at this size the
cloze results are the meaningful ones. Chance is 25%.

### Base vs instruct

Both are scored on the same raw-text prompts, without the chat template, so the numbers are directly comparable. SFT mainly teaches format, turn-taking and when to stop, so large changes in knowledge benchmarks aren't expected. Chance on the multiple-choice tasks is 25%.

| Metric | Base | **Instruct** |
|---|---|---|
| SciQ (acc_norm) | 53.9% | 55.3% |
| ARC-Easy (acc_norm) | 32.4% | 32.2% |
| HellaSwag (acc_norm) | 26.3% | 26.7% |
| MMLU (letter) | 25.2% | 25.0% |
| MMLU (cloze) | 25.7% | 25.9% |
| MMLU biology & medicine (cloze) | 27.8% | 27.6% |

### Compared with its siblings and other small models

| | **This model** | `tiny` | [Pythia-70M](https://huggingface.co/EleutherAI/pythia-70m) | [GPT-2](https://huggingface.co/openai-community/gpt2) | [Supra-50M](https://huggingface.co/SupraLabs/Supra-50M-Base) |
|---|---|---|---|---|---|
| Parameters (total) | 3.8M | 2.0M | 70.4M | 124.4M | 51.8M |
| Active non-embedding parameters | 0.5M | 0.5M | 18.9M | 85.1M | 35.4M |
| Training tokens | 1,000M | 1,000M | 300B (the Pile) | undisclosed (WebText) | 20B (FineWeb-Edu) |
| SciQ (acc_norm) | 55.3% | 56.3% | 55.2% | 53.2% | 77.2% |
| ARC-Easy (acc_norm) | 32.2% | 32.7% | 35.0% | 42.0% | 52.2% |
| HellaSwag (acc_norm) | 26.7% | 26.7% | – | 29.5% | 31.8% |
| MMLU (cloze) | 25.9% | 25.8% | – | – | – |
| MMLU biology & medicine (cloze) | 27.6% | 28.4% | – | – | – |
| Validation loss (held-out mix) | – | 3.869 | – | – | – |

LittleKedi columns are measured with this repo's `evals.py` and `mmlu_eval.py` on the checkpoints named. Validation
losses are only comparable between models that share a tokenizer and validation set. The other columns are
**published figures**: Pythia-70M from EleutherAI's evaluations as quoted on the
[Wisp-15M card](https://huggingface.co/DedeProGames/Wisp-15M), GPT-2 and Supra-50M from the
[Supra-50M-Base card](https://huggingface.co/SupraLabs/Supra-50M-Base). Their prompts may differ from ours,
so treat gaps of a few points as noise.

## What it can and can't do

This is the instruction-tuned **sparsest tiny model**: the same 0.5M active non-embedding parameters per token
as `tiny-instruct`, with ~3× the expert capacity. Fine-tuning taught it the shape of a conversation, and it
fits the chat data better than tiny (held-out SFT loss 3.15 against 3.30), but it's still far too small to be an
assistant. It's a research artifact for studying sparsity and instruction tuning at the smallest scale.

**What it does**

- ✅ Replies in the chat format: a short paragraph that usually ends with `<|end|>`.
- ✅ Stays on the topic of the question more often than `tiny-instruct`: *"What's a good name for a cat?"* →
  *"A good name for a cat is…"*; asked about exam stress, it talks about sleep and ends *"Good luck…"*; asked how
  radio changed news, it talks about media and technology.
- ✅ Writes grammatical English in the register of an answer.

**What it can't do**

- ❌ **Give correct answers.** On-topic replies are still invented: *"What causes the seasons on Earth?"* →
  *"Air-directed volcanoes are critical to the long-term impact of the POW earthquake."*
- ❌ **Facts**, even about text from its own pretraining data. Treat every name, date and number as made up.
- ❌ **Arithmetic.** Asked *"What is 7 + 6?"* it fell into a sub-word loop (`ahvahvahv…`) and never stopped.
  The repetition penalty doesn't catch loops of word fragments, so cap the reply length.
- ❌ **Creative requests, instructions, code.**

**Benchmarks:** about where the base model was (SciQ 55.3%, ARC-Easy 32.2%, MMLU cloze 25.9%): instruction
tuning changes the format, not what the model knows.

**Decoding:** temperature ≈ 0.7, top-p ≈ 0.9, repetition penalty ≈ 1.15 (the GGUF files carry these as
defaults), and a modest maximum length.

## Usage

**llama.cpp** (GGUF, llama.cpp's built-in `deepseek2` architecture; the chat template is embedded):

```sh
llama-cli -m LittleKedi-tiny-wide-sparse-instruct-q8_0.gguf -cnv --temp 0.7 --top-p 0.9 --repeat-penalty 1.15
```

**Ollama** (uses the bundled `template` and `params` files): `ollama run hf.co/edededdy/LittleKedi-tiny-wide-sparse-instruct:Q8_0`

**PyTorch** (needs the `littlekedi` package from [the training code](https://github.com/EdmundMartin/LittleKedi)):

```python
import json, torch
from safetensors.torch import load_file
from littlekedi import LittleKediModel, ModelConfig
from littlekedi.chat import END, render_prompt
from littlekedi.tokenizer import BPETokenizer

model = LittleKediModel(ModelConfig.from_dict(json.load(open("config.json")))).eval()
model.load_state_dict(load_file("model.safetensors"))
tok = BPETokenizer.from_file("tokenizer.json")
ids = torch.tensor([render_prompt(tok, [{"role": "user", "content": "What causes the seasons?"}])])
out = model.generate(ids, 200, temperature=0.7, top_p=0.9, repetition_penalty=1.15, stop_ids={tok.token_to_id(END)})
print(tok.decode(out[0, ids.shape[1]:].tolist()).replace(END, "").strip())
```

Or in a terminal: `python chat.py --ckpt <checkpoint>.pt`.

## Files

| File | Contents |
|---|---|
| `model.safetensors` | PyTorch weights (embedding tied with the output head), `littlekedi` parameter names |
| `config.json` | `ModelConfig` for `littlekedi` |
| `tokenizer.json` | Byte-level BPE tokenizer (Hugging Face `tokenizers` format) |
| `chat_template.jinja` | Chat template (Jinja), the same one embedded in the GGUF files |
| `tokenizer_config.json` | Special tokens (`<|end|>` = end of turn), chat template, context length |
| `template` | Ollama chat template |
| `params` | Ollama stop tokens and sampling defaults |
| `LittleKedi-tiny-wide-sparse-instruct-f16.gguf` | GGUF (f16) for llama.cpp / Ollama / LM Studio |
| `LittleKedi-tiny-wide-sparse-instruct-q8_0.gguf` | GGUF (q8_0) for llama.cpp / Ollama / LM Studio |
| `little_kedi.jpg` | The LittleKedi logo |

## Acknowledgements

- The architecture follows **DeepSeek-V3** (DeepSeek-AI, 2024): MLA, DeepSeekMoE and auxiliary-loss-free balancing.
- **Muon** (Keller Jordan et al., 2024) with Moonlight's update scaling (Liu et al., 2025).
- Pretraining data: **FineWeb-Edu** (ODC-BY 1.0).
- Pretraining data: **Cosmopedia** (Apache 2.0).
- Pretraining data: **FineMath 4+** (ODC-BY 1.0).
- Pretraining data: **SmolTalk** (Apache 2.0 for its new subsets, with the datasets it includes under their own licenses).
- Pretraining data: **github-code-clean** (Apache 2.0, with each file under its own open-source license).
- Chat data: see the SFT table above for its sources, each under its own license.

*Card generated 2026-10-13 by `prepare_release.py` from `tiny-wide-sparse-sft.pt` (step 15257).*