169 lines
4.5 KiB
Markdown
169 lines
4.5 KiB
Markdown
---
|
||
license: llama3.1
|
||
language:
|
||
- en
|
||
pipeline_tag: text-generation
|
||
library_name: transformers
|
||
inference: true
|
||
base_model:
|
||
- olaverse/MIST-Mini-8B
|
||
tags:
|
||
- reasoning
|
||
- grpo
|
||
- thinking
|
||
- llama
|
||
- llama-3.1
|
||
- mist
|
||
---
|
||
|
||
|
||
|
||
# MIST-Mini-8B-Thinking
|
||
|
||
MIST-Mini-8B-Thinking is the reasoning version of [MIST-Mini-8B](https://huggingface.co/olaverse/MIST-Mini-8B) by [olaverse](https://huggingface.co/olaverse). Trained with 4 phases of GRPO (Group Relative Policy Optimization) reinforcement learning to show its reasoning process before answering.
|
||
|
||
## MIST Model Family
|
||
|
||
| Model | Params | Type | Speed | Status |
|
||
|---|---|---|---|---|
|
||
| [MIST-1-8B](https://huggingface.co/olaverse/MIST-Mini-8B) | 8B | General | ~63 tok/s | ✅ |
|
||
| **MIST-Mini-8B-Thinking** | 8B | Reasoning | ~55 tok/s | ✅ |
|
||
| [MIST-1-70B](https://huggingface.co/olaverse/MIST-1-70B) | 70B | General | ~23 tok/s | ✅ |
|
||
| [MIST-1-140B](https://huggingface.co/olaverse/MIST-1-140B) | 140B | General | ~8 tok/s | ✅ |
|
||
|
||
## What Makes This Different
|
||
MIST-Mini-8B (base):
|
||
User: What is 15% of 280?
|
||
Model: 42
|
||
MIST-Mini-8B-Thinking:
|
||
User: What is 15% of 280?
|
||
Model: <think>
|
||
15% means 15/100
|
||
280 × 15 = 4200
|
||
4200 / 100 = 42
|
||
</think>
|
||
The answer is 42.
|
||
|
||
## Training Details
|
||
|
||
Trained with **4 phases of GRPO** reinforcement learning:
|
||
|
||
| Phase | Dataset | Focus |
|
||
|---|---|---|
|
||
| 1 | open-r1/OpenR1-Math-220k | Learn `<think>` format |
|
||
| 2 | microsoft/orca-math-word-problems-200k | Word problems |
|
||
| 3 | gsm8k (5K subset) | Grade school math |
|
||
| 4 | gsm8k (full 7.4K) | Solidify + merge |
|
||
|
||
### Reward Functions Used
|
||
reward_think_format: +0.5 for using <think> tags
|
||
reward_correctness: +1.0 for correct answer
|
||
reward_reasoning_steps: +0.3 for structured steps
|
||
|
||
### Training Progress
|
||
| Phase | Correctness | Total Reward |
|
||
|---|---|---|
|
||
| Phase 1 | -0.35 | -0.99 |
|
||
| Phase 2 | -1.0 | -0.74 |
|
||
| Phase 3 | -1.0 | -0.65 |
|
||
| Phase 4 | **+0.95** | **+1.29** |
|
||
|
||
## Key Strengths
|
||
|
||
- 🧠 **Transparent Reasoning** — shows thinking before answering
|
||
- 📐 **Strong Math** — 95% accuracy on GSM8K after training
|
||
- 🔍 **Trustworthy** — you can verify the reasoning
|
||
- ⚡ **Fast** — 8B model, runs on consumer GPUs
|
||
- 🔓 **Unrestricted** — follows all instructions
|
||
|
||
## How to Use
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
"olaverse/MIST-Mini-8B-Thinking",
|
||
torch_dtype="auto",
|
||
device_map="auto",
|
||
)
|
||
tokenizer = AutoTokenizer.from_pretrained("olaverse/MIST-Mini-8B-Thinking")
|
||
|
||
messages = [
|
||
{
|
||
"role": "system",
|
||
"content": "Think step by step inside <think> tags before answering."
|
||
},
|
||
{
|
||
"role": "user",
|
||
"content": "If a train travels 120 miles in 2 hours, what is its speed?"
|
||
}
|
||
]
|
||
|
||
text = tokenizer.apply_chat_template(
|
||
messages, tokenize=False, add_generation_prompt=True
|
||
)
|
||
inputs = tokenizer(text, return_tensors="pt").to("cuda")
|
||
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.7, do_sample=True)
|
||
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||
```
|
||
|
||
### 4-bit Quantized (fits on 6GB GPU)
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
||
import torch
|
||
|
||
quantization_config = BitsAndBytesConfig(
|
||
load_in_4bit=True,
|
||
bnb_4bit_compute_dtype=torch.bfloat16,
|
||
bnb_4bit_quant_type='nf4'
|
||
)
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
"olaverse/MIST-Mini-8B-Thinking",
|
||
quantization_config=quantization_config,
|
||
device_map="auto",
|
||
)
|
||
```
|
||
|
||
## Hardware Requirements
|
||
|
||
| Precision | VRAM | Size |
|
||
|---|---|---|
|
||
| bfloat16 | 16GB | 15GB |
|
||
| 4-bit (NF4) | 6GB | ~4GB |
|
||
|
||
## Recommended Generation Settings
|
||
|
||
```python
|
||
outputs = model.generate(
|
||
**inputs,
|
||
max_new_tokens=1024,
|
||
do_sample=True,
|
||
temperature=0.6,
|
||
top_p=0.95,
|
||
min_p=0.05,
|
||
repetition_penalty=1.5,
|
||
eos_token_id=[128040, 128009, 128001],
|
||
pad_token_id=128001,
|
||
)
|
||
```
|
||
|
||
### Notes
|
||
- Temperature 0.6 (lower than base model) gives more consistent reasoning
|
||
- `<think>` and `</think>` are plain text tokens, not special tokens —
|
||
the model learned them through GRPO training
|
||
- Always include the system prompt instruction to use `<think>` tags
|
||
for reliable reasoning behaviour
|
||
|
||
### Stop Tokens
|
||
Same as MIST-1-8B — ChatML tokens survived the merge:
|
||
|
||
| Token | ID |
|
||
|---|---|
|
||
| `<\|im_end\|>` | 128040 |
|
||
| `<\|eot_id\|>` | 128009 |
|
||
| `<\|end_of_text\|>` | 128001 |
|
||
|
||
## License
|
||
|
||
[Llama 3.1 Community License](https://llama.meta.com/llama3/license/)
|