99 lines
2.8 KiB
Markdown
99 lines
2.8 KiB
Markdown
|
---
|
||
|
|
license: agpl-3.0
|
||
|
|
base_model: unsloth/DeepSeek-R1-0528-Qwen3-8B
|
||
|
|
tags:
|
||
|
|
- marxism-leninism
|
||
|
|
- grpo
|
||
|
|
- llama-cpp
|
||
|
|
- ollama
|
||
|
|
- political-education
|
||
|
|
- marxism
|
||
|
|
- communism
|
||
|
|
- political-extremism
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
# MLMLML - Machine Learning Marxist-Leninist Models of Language
|
||
|
|
|
||
|
|
A GRPO fine-tuned language model for Marxist-Leninist political education and analysis.
|
||
|
|
|
||
|
|
## Model Description
|
||
|
|
|
||
|
|
This model is fine-tuned from `unsloth/DeepSeek-R1-0528-Qwen3-8B` using Group Relative Policy Optimization (GRPO)
|
||
|
|
on a curated dataset of Marxist-Leninist Q&A pairs from [ProleWiki](https://en.prolewiki.org/).
|
||
|
|
|
||
|
|
The training rewards:
|
||
|
|
- **Ideological firmness**: Clear positions grounded in material analysis
|
||
|
|
- **Coherence**: Self-consistent, well-structured responses
|
||
|
|
- **Accuracy**: Faithful to Marxist-Leninist theory and historical evidence
|
||
|
|
|
||
|
|
The training penalizes:
|
||
|
|
- "Both-sidesing" and false balance
|
||
|
|
- Hedging and evasive language
|
||
|
|
- Bourgeois framing and ahistorical claims
|
||
|
|
|
||
|
|
## Writing Style
|
||
|
|
|
||
|
|
Following Chairman Mao's guidance in "Oppose Stereotyped Party Writing":
|
||
|
|
- **Vigorous, lively, fresh and forceful** - never drab or stereotyped
|
||
|
|
- **Audience-aware** - "When shooting an arrow, one must aim at the target"
|
||
|
|
- **Investigation-based** - "No investigation, no right to speak"
|
||
|
|
- **Clear positions** - FOR and AGAINST, using scientific argument
|
||
|
|
|
||
|
|
## Usage
|
||
|
|
|
||
|
|
### Download and Convert to GGUF
|
||
|
|
|
||
|
|
```bash
|
||
|
|
# Clone the repo
|
||
|
|
git lfs install
|
||
|
|
git clone https://huggingface.co/percyraskova/MLMLML
|
||
|
|
cd MLMLML
|
||
|
|
|
||
|
|
# Convert to GGUF (requires llama.cpp)
|
||
|
|
python ~/llama.cpp/convert_hf_to_gguf.py . --outfile MLMLML-F16.gguf --outtype f16
|
||
|
|
|
||
|
|
# Quantize to Q4_K_M
|
||
|
|
~/llama.cpp/build/bin/llama-quantize MLMLML-F16.gguf MLMLML-Q4_K_M.gguf Q4_K_M
|
||
|
|
|
||
|
|
# Create Ollama model
|
||
|
|
ollama create mlmlml -f Modelfile
|
||
|
|
ollama run mlmlml
|
||
|
|
```
|
||
|
|
|
||
|
|
### Direct with Transformers
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
|
||
|
|
model = AutoModelForCausalLM.from_pretrained("percyraskova/MLMLML")
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained("percyraskova/MLMLML")
|
||
|
|
|
||
|
|
inputs = tokenizer("What is imperialism?", return_tensors="pt")
|
||
|
|
outputs = model.generate(**inputs, max_new_tokens=512)
|
||
|
|
print(tokenizer.decode(outputs[0]))
|
||
|
|
```
|
||
|
|
|
||
|
|
## Training Details
|
||
|
|
|
||
|
|
- **Base model**: unsloth/DeepSeek-R1-0528-Qwen3-8B
|
||
|
|
- **Method**: GRPO (Group Relative Policy Optimization)
|
||
|
|
- **Dataset**: ProleWiki Q&A pairs (~4500 samples)
|
||
|
|
- **Epochs**: 2
|
||
|
|
- **Hardware**: NVIDIA A100 80GB
|
||
|
|
|
||
|
|
## Limitations
|
||
|
|
|
||
|
|
This model is designed for educational purposes about Marxist-Leninist theory and analysis.
|
||
|
|
It takes clear ideological positions and is not intended to be "neutral" on class struggle,
|
||
|
|
imperialism, or other questions where Marxism-Leninism has definite answers.
|
||
|
|
|
||
|
|
## License
|
||
|
|
|
||
|
|
Apache 2.0
|
||
|
|
|
||
|
|
## Citation
|
||
|
|
|
||
|
|
If you use this model, please cite ProleWiki as the source of training data.
|