---
language:
- ru
- en
license: apache-2.0
base_model: Rahspide/MNN-Pocket-0.3B-RU
library_name: transformers
pipeline_tag: text-generation
datasets:
- IlyaGusev/saiga_scored
- HuggingFaceH4/ultrachat_200k
tags:
- russian
- chat
- assistant
- on-device
- mnn
- qwen3
- gguf
- work-in-progress
---

![MNN-Pocket-0.3B-RU-Chat — a Russian assistant in 0.3B](assets/cover_chat.png)

![params](https://img.shields.io/badge/params-361M-E2543F?style=flat-square)
![sft](https://img.shields.io/badge/sft-50M%20tokens-5FD3B4?style=flat-square)
![data](https://img.shields.io/badge/data-18.8M%20ru%20%2B%206M%20en-6FA8FF?style=flat-square)
![gguf](https://img.shields.io/badge/gguf-Q8__0%20%C2%B7%20Q4__K__M-E2543F?style=flat-square)
![license](https://img.shields.io/badge/license-apache--2.0-8B95A4?style=flat-square)
![status](https://img.shields.io/badge/status-v1%20released-5FD3B4?style=flat-square)

> **Status: v1 released.** Instruction-tuned on 50 M tokens (18.8 M Russian + 6 M English replay) on one Kaggle
> T4 in ~6.5 h. Weights and GGUF builds are in this repository.

Russian instruction-following assistant, built by SFT on top of
[`Rahspide/MNN-Pocket-0.3B-RU`](https://huggingface.co/Rahspide/MNN-Pocket-0.3B-RU). It answers instead of
continuing; that dialogue shape is what 50 M tokens of SFT bought.

## Two models, two jobs

![One base, two jobs](assets/two_jobs_chat.png)

| | Sister — `MNN-Pocket-0.3B-RU` | This model — `MNN-Pocket-0.3B-RU-Chat` |
|---|---|---|
| Job | continue text | follow instructions |
| Input | a prefix, often mid-sentence | a system + user message |
| Output | the plausible continuation | an answer |
| Trained with | continued pretraining on 1.2 B tokens of raw text | instruction tuning on top of that model |
| Choose it for | keyboard suggestions, autocomplete, prefix completion | chat, Q&A, short rewrites |
| Do not choose it for | assistant behaviour: it will continue your question instead of answering it | autocomplete: it will answer instead of continuing |

The two tasks reward opposite behaviour. Trained on both jobs, one checkpoint does each badly: it answers when
it should continue, and continues when it should answer.

## Why start from the Russian continuation model

The base `MNN-Pocket-0.3B` is bilingual zh/en and has essentially no Russian; a direct SFT of the official
instruct model would fight both problems at once (new language and new format). Starting from the continuation
model means Russian is already paid for, and SFT only has to teach the dialogue shape.

## Recipe (as run)

- Start point: `Rahspide/MNN-Pocket-0.3B-RU` (step 36,621).
- Chat template: the model's own ChatML (`<|im_start|>` / `<|im_end|>`), untouched.
- Loss only on assistant tokens; system and user turns are masked context.
- Multi-turn dialogue kept as one example; packing with explicit boundaries.
- One run at 50 M tokens (1,525 optimizer steps, 32 k tokens each), checkpoints every 300 steps.
- 85 / 15 Russian-to-English sampling per micro-batch; the English share keeps bilingual instruction following.
- lr 1e-4, cosine, 5% warmup, full fine-tuning, fp16 autocast on T4 with fp32 master weights.
- ~6.5 h on one Kaggle T4.

## Data (as used)

![Training mix](assets/mix.png)

| Dataset | Role | Licence |
|---|---|---|
| [`IlyaGusev/saiga_scored`](https://huggingface.co/datasets/IlyaGusev/saiga_scored) | core Russian instructions, 18.8 M tokens / 23,518 dialogues | see dataset card |
| [`HuggingFaceH4/ultrachat_200k`](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) | English replay, 6.0 M tokens / 5,181 dialogues | MIT |

Filters applied to `saiga_scored`: Russian only, dropped rows flagged `is_bad_by_regex`, 80–6,000 characters per
dialogue, 2–8 turns. Roles `bot` / `gpt` are normalised to `assistant`; empty turns are dropped.

## Quality spot-check

Greedy generation, Q4_K_M, CPU:

| Prompt | Answer |
|---|---|
| «Объясни в двух предложениях, что такое инфляция.» | «Инфляция — это процесс роста цен на товары и услуги в стране или за ее пределами. Этот процесс может привести к у…» |
| «Напиши короткое поздравление с днём рождения коллеге.» | «Дорогой [Имя], С днём рождения! Желаю, чтобы ваша работа была не только эффективной, но и приятной для глаз…» |
| «Привет! Как дела?» | repeats the question — a known weak spot of 0.3 B on very short turns |

The model answers instead of continuing the prompt. That is all this table proves.

## Builds

![GGUF sizes](assets/sizes.png)

| Build | Size | Download | Note |
|---|---|---|---|
| Q8_0 | 310 MB | [`MNN-Pocket-0.3B-RU-Chat-Q8_0.gguf`](https://huggingface.co/Rahspide/MNN-Pocket-0.3B-RU-Chat/blob/main/gguf/MNN-Pocket-0.3B-RU-Chat-Q8_0.gguf) | near-lossless |
| Q4_K_M | 223 MB | [`MNN-Pocket-0.3B-RU-Chat-Q4_K_M.gguf`](https://huggingface.co/Rahspide/MNN-Pocket-0.3B-RU-Chat/blob/main/gguf/MNN-Pocket-0.3B-RU-Chat-Q4_K_M.gguf) | phone size |
| MNN INT4 | 165 MB | [mnn/MNN-Pocket-0.3B-RU-Chat-INT4](https://huggingface.co/Rahspide/MNN-Pocket-0.3B-RU-Chat/tree/main/mnn/MNN-Pocket-0.3B-RU-Chat-INT4) | Android, MNN runtime |

```bash
llama-cli -hf Rahspide/MNN-Pocket-0.3B-RU-Chat:Q4_K_M --jinja -p "Объясни, что такое инфляция" -n 60
```

The MNN build runs on MNN's runtime (`llm_demo`, built with `-DMNN_BUILD_LLM=ON`). Its `llm_config.json` keeps
the ChatML template, so it answers out of the box:

```bash
cd mnn/MNN-Pocket-0.3B-RU-Chat-INT4 && llm_demo config.json prompt.txt
```

165 MB on disk, ~497 tok/s prefill and ~80 tok/s decode on 12 CPU threads. The answer is coherent Russian; the
INT4 weights cost some precision against Q8_0, as expected.

**~123 tok/s generation on Q4_K_M, ~102 tok/s on Q8_0** (16-thread x86 CPU, no GPU). Apply the model's own ChatML
template (`--jinja` in llama.cpp, `apply_chat_template` in transformers). Without it the model continues text
instead of answering and behaves like its sister.

## Reproduce

```bash
# chat smoke test through the model's own template
python bench/chat_smoke.py --model Rahspide/MNN-Pocket-0.3B-RU-Chat --cpu

# throughput of the quantised build
LLAMA_CLI=/path/to/llama-cli bash bench/bench_cpu.sh gguf/MNN-Pocket-0.3B-RU-Chat-Q4_K_M.gguf --chat
```

`bench/chat_smoke.py` applies the chat template from `tokenizer_config.json`, generates greedily and asserts that
the answer does not echo the prompt, which is a cheap check that the template and the checkpoint still match.
`bench/bench_cpu.sh --chat` wraps the prompt in ChatML before timing, so the numbers are comparable with the
table above.

## Evaluation

60 held-out Russian prompts from the `saiga_scored` validation slice, greedy decoding, the model's own template:

| Automatic check | Result |
|---|---|
| answers in Russian (>50% Cyrillic) | 57 / 60 |
| mean Cyrillic share of letters | 0.93 |
| empty answers | 0 |
| echoes the prompt | 2 |
| mean answer length, words | 26.1 |
| distinct-2 (higher = less repetition) | 0.90 |

The sheet with all 60 prompts and answers is in
[`review/blind_sheet.md`](https://huggingface.co/Rahspide/MNN-Pocket-0.3B-RU-Chat/blob/main/review/blind_sheet.md).

A comparison against `taobao-mnn/MNN-Pocket-0.3B` was attempted on the same prompts and dropped. That checkpoint
degenerates on Russian input, repeating the prompt in bold markdown fragments:

```
Q: Как стать популярным в яндекс дзен?
A: Конечный откорт: **"Популярный в яндекс дзен"** — это **не** в яндекс (в яндекс) — **в яндекс** — **в яндекс** — …
```

So it is not a baseline.

## Limitations

- 0.36 B parameters: short answers, weak facts, no reliable multi-step reasoning.
- Russian SFT data is mostly machine-translated or model-generated; style is flatter than human chat.
- The sister's tokenizer cost applies here too: about 3.3 tokens per Russian word, so ~3× more decode steps per
  Russian sentence than for English. The vocabulary is byte-level, so rare and uppercase Cyrillic letters can
  end up split across tokens; a streaming decoder has to buffer partial UTF-8 sequences.
- Not a coding model, not a tool-calling model, no long-context dialogue.
- English answers are thinner than Russian ones. Replay keeps the ability alive; it does not polish it.

## Releases

| Date | Version | What changed |
|---|---|---|
| 2026-11-10 | **v1.0** | Assistant released: 50 M-token SFT (18.8 M ru + 6 M en), bf16 weights, GGUF Q8_0 / Q4_K_M, ChatML smoke test passing |
| 2026-11-11 | v1.1 | MNN INT4 build for Android: 165 MB, ChatML template included, ~80 tok/s decode |
| 2026-11-10 | v1.0 (sister) | Continuation checkpoint released: 1.2 B-token CPT, same base |

## Licence and attribution

Weights: Apache-2.0, inherited through `MNN-Pocket-0.3B-RU` from `taobao-mnn/MNN-Pocket-0.3B-Base`.

```bibtex
@misc{mnn_pocket_03b,
  title  = {MNN-Pocket-0.3B},
  author = {Taobao MNN team},
  year   = {2026},
  url    = {https://huggingface.co/taobao-mnn/MNN-Pocket-0.3B}
}
```