179 lines
6.5 KiB
Markdown
179 lines
6.5 KiB
Markdown
---
|
||
language:
|
||
- en
|
||
license: apache-2.0
|
||
library_name: transformers
|
||
pipeline_tag: text-generation
|
||
tags:
|
||
- moe
|
||
- mixture-of-experts
|
||
- chat
|
||
- sft
|
||
- fine-tuned
|
||
- pytorch
|
||
- tiny
|
||
- efficient
|
||
- trl
|
||
base_model: FlameF0X/TinyMoE-100m-2x8-retrained
|
||
datasets:
|
||
- HuggingFaceTB/smol-smoltalk
|
||
- yahma/alpaca-cleaned
|
||
- databricks/databricks-dolly-15k
|
||
- HuggingFaceH4/ultrachat_200k
|
||
- Open-Orca/OpenOrca
|
||
- teknium/openhermes
|
||
- HuggingFaceH4/no_robots
|
||
model-index:
|
||
- name: TinyMoE-100m-2x8-chat-stage1
|
||
results: []
|
||
---
|
||
|
||
# TinyMoE-100m-2x8 Chat (Stage 1)
|
||
|
||
A **chat fine-tuned Mixture of Experts (MoE) language model** — small, efficient, and fully open-source.
|
||
|
||
TinyMoE is a ~100M parameter transformer using a **Mixture of Experts** architecture, fine-tuned via supervised fine-tuning (SFT) on a diverse blend of high-quality chat and instruction datasets. It's designed to be a compact but capable conversational AI that knows its own identity.
|
||
|
||
## Model Details
|
||
|
||
| Property | Value |
|
||
|---|---|
|
||
| **Architecture** | Mixture of Experts (MoE) Transformer |
|
||
| **Parameters** | ~100M total |
|
||
| **Experts** | 8 experts, top-2 routing per token |
|
||
| **Context Length** | 4,096 tokens (extended via linear RoPE scaling) |
|
||
| **Chat Template** | ChatML-style — `<\|system\|>`, `<\|user\|>`, `<\|assistant\|>` |
|
||
| **Base Model** | [TinyMoE-100m-2x8-retrained](https://huggingface.co/FlameF0X/TinyMoE-100m-2x8-retrained) |
|
||
| **Training Type** | Full-weight SFT (not LoRA) |
|
||
| **Precision** | bfloat16 |
|
||
| **Creator** | [FlameF0X](https://huggingface.co/FlameF0X) |
|
||
| **License** | Apache 2.0 |
|
||
|
||
## What is Mixture of Experts?
|
||
|
||
Unlike a standard dense transformer where every token goes through the same large feed-forward network, TinyMoE uses **8 expert sub-networks** with a learned router that selects the top-2 experts per token. This means:
|
||
|
||
- **More total knowledge capacity** without proportionally increasing compute
|
||
- **Sparse activation** — only a fraction of parameters fire per token
|
||
- **Efficient inference** — you get more model per FLOP
|
||
|
||
This is the same architectural family as Mixtral, but at a much smaller scale — proving that MoE works even at ~100M parameters.
|
||
|
||
## Training Recipe
|
||
|
||
### Stage 1 — Chat SFT
|
||
|
||
The base pretrained model was fine-tuned on a carefully curated mixture of chat datasets to teach conversational ability, instruction following, and model identity.
|
||
|
||
**Hardware:** NVIDIA L4 (24 GB) on [Modal](https://modal.com)
|
||
**Framework:** TRL (Transformer Reinforcement Learning) `SFTTrainer`
|
||
**Max examples:** 80,000 (after length filtering)
|
||
|
||
| Hyperparameter | Value |
|
||
|---|---|
|
||
| Epochs | 3 |
|
||
| Learning Rate | 2e-5 |
|
||
| LR Schedule | Cosine with 5% warmup |
|
||
| Optimizer | AdamW (weight decay 0.01) |
|
||
| Batch Size | 4 per device × 8 grad accum = 32 effective |
|
||
| Max Gradient Norm | 1.0 |
|
||
| Packing | Yes |
|
||
| Gradient Checkpointing | Yes |
|
||
|
||
### Training Datasets
|
||
|
||
| Dataset | Examples | Description |
|
||
|---|---|---|
|
||
| [SmolTalk](https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk) | ~4k | Synthetic diverse chat conversations |
|
||
| [Alpaca Cleaned](https://huggingface.co/datasets/yahma/alpaca-cleaned) | ~52k | Cleaned instruction-following data |
|
||
| [Dolly 15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) | ~15k | Human-written instruction/response pairs |
|
||
| [UltraChat 200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) | 12k subset | High-quality multi-turn chat |
|
||
| [OpenOrca](https://huggingface.co/datasets/Open-Orca/OpenOrca) | 15k subset | GPT-4 augmented FLAN instructions |
|
||
| [OpenHermes](https://huggingface.co/datasets/teknium/openhermes) | 10k subset | Diverse tasks (code, write, reason, roleplay) |
|
||
| [No Robots](https://huggingface.co/datasets/HuggingFaceH4/no_robots) | 10k | Hand-curated high-quality SFT examples |
|
||
| **TinyMoE Identity** *(custom)* | 44 | Synthetic identity Q&A + multi-turn conversations |
|
||
|
||
### Identity Training
|
||
|
||
The model was explicitly taught to know it's **TinyMoE** through two mechanisms:
|
||
|
||
1. **Dedicated identity dataset** — 44 custom examples covering name, creator, architecture, capabilities, and differentiation from other models (ChatGPT, Claude, Llama, etc.)
|
||
2. **System prompt injection** — 15% of all training examples received a system prompt: *"You are TinyMoE, a helpful Mixture of Experts AI assistant created by FlameF0X."*
|
||
|
||
This means TinyMoE knows who it is — ask it "What's your name?" or "Who created you?" and it will answer correctly.
|
||
|
||
## Usage
|
||
|
||
### Chat Format
|
||
|
||
TinyMoE uses a **ChatML-style** template with special tokens:
|
||
|
||
```
|
||
<|system|>
|
||
You are TinyMoE, a helpful Mixture of Experts AI assistant created by FlameF0X.</s>
|
||
<|user|>
|
||
What's Mixture of Experts?</s>
|
||
<|assistant|>
|
||
MoE stands for Mixture of Experts! Instead of one big neural network...</s>
|
||
```
|
||
|
||
### Quick Start
|
||
|
||
```python
|
||
import torch
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model_id = "FlameF0X/TinyMoE-100m-2x8-chat-stage1"
|
||
|
||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_id,
|
||
torch_dtype=torch.bfloat16,
|
||
device_map="auto",
|
||
)
|
||
|
||
messages = [
|
||
{"role": "system", "content": "You are TinyMoE, a helpful Mixture of Experts AI assistant created by FlameF0X."},
|
||
{"role": "user", "content": "What's your name and who made you?"},
|
||
]
|
||
|
||
inputs = tokenizer.apply_chat_template(
|
||
messages,
|
||
tokenize=True,
|
||
add_generation_prompt=True,
|
||
return_tensors="pt",
|
||
).to(model.device)
|
||
|
||
outputs = model.generate(
|
||
inputs,
|
||
max_new_tokens=256,
|
||
temperature=0.7,
|
||
do_sample=True,
|
||
top_p=0.9,
|
||
)
|
||
|
||
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||
```
|
||
## Capabilities & Limitations
|
||
|
||
### ✅ Strengths
|
||
|
||
- **Efficient** — MoE architecture means more capacity per inference FLOP
|
||
- **Conversational** — trained on diverse multi-turn chat data
|
||
- **Self-aware** — knows it's TinyMoE, not ChatGPT/Claude/Llama
|
||
- **Open-source** — weights, architecture, and training code are all public
|
||
- **Fast** — small enough to run on consumer hardware or free-tier GPUs
|
||
|
||
### ⚠️ Limitations
|
||
|
||
- **Small model** — at 100M parameters, factual knowledge is limited compared to billion-parameter models
|
||
- **Stage 1 only** — this is a direct-answer SFT model; it hasn't undergone RLHF/DPO alignment
|
||
- **No CoT** — training explicitly excluded chain-of-thought reasoning traces (saved for future stages)
|
||
- **English only** — training data was English-dominant
|
||
- **May hallucinate** — like all LLMs, it can generate incorrect information with confidence
|
||
|
||
### Future Stages (planned)
|
||
|
||
- **Stage 2:** MCP/Tool usage + RL
|
||
- **Longer context** — potential extension beyond 4K tokens
|