salakash/SamKash-Tolstoy

Model Card for Model ID

SamKash-Tolstoy — DeepSeek LoRA (Russian Literature-Fine-tuned for English-language content only)

https://huggingface.co/blog/salakash/cpu-only-micro-llm-tolstoy

Developed by Kashif Salahuddin and Samiya Kashif, SamKash-Tolstoy is a domain-specialized LLM (lightweight LoRA adapter) built exclusively for Russian literature (English-only). It’s trained on 475 public-domain Russian classics from the Project Gutenberg collection and enriched with university and critics’ articles filtered from the OSCAR web corpus, so the voice and psychological depth feel authentic without using any copyrighted books.

Reasoning-forward core: Based on deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B, giving strong structure and long-form coherence; further supervised fine-tuning from output feedback reduces drift and hallucinations over time.

Canon-focused: Tolstoy, Dostoevsky, Turgenev, Chekhov, Gogol, and peers—curated for style, theme, and historical register.

Ethically sourced: Only no-copyright texts; web articles filtered for relevance to Russian literature.

Built for creators & scholars: Draft scenes and monologues, analyze motifs, outline lectures, or explore stylistic transformations—fast.

Hugging Face Repo: salakash/SamKash-Tolstoy

Example prompt: “Write a short scene in the style of Crime and Punishment: a feverish student crosses a Petersburg bridge at night.”

---

TL;DR: Use It

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
from peft import PeftModel

base_id = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
adpt_id = "salakash/SamKash-Tolstoy"   # replace with your repo path

tok = AutoTokenizer.from_pretrained(base_id, use_fast=True)

# CPU (float32) or Apple M-series (MPS, float16)
import torch
device = "mps" if torch.backends.mps.is_available() else "cpu"
dtype   = torch.float16 if device == "mps" else torch.float32

base = AutoModelForCausalLM.from_pretrained(base_id, dtype=dtype)
base.to(device)

model = PeftModel.from_pretrained(base, adpt_id)
model.config.use_cache = True  # inference = OK to re-enable KV cache

gen = pipeline("text-generation", model=model, tokenizer=tok, device=-1)
out = gen(
    "Write a reflective paragraph about conscience and fate in an aristocratic household.",
    max_new_tokens=200, do_sample=True, temperature=0.7, top_p=0.9
)[0]["generated_text"]
print(out)




## Model Details

### Model Description

- **Developed by:** Samiya Kashif & Kashif Salahuddin
- **Funded by:** Self-funded (individual project)
- **Shared by:** Samiya Kashif & Kashif Salahuddin
- **Model type:** LoRA (PEFT) adapter for a decoder-only causal language model (Qwen-family, 1.5B params base)
- **Language(s) (NLP):** English
- **License:** Apache 2.0
- **Finetuned from model:** deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

### Model Sources [optional]

<!-- Provide the basic links for the model. -->

- **Repository:** salakash/SamKash-Tolstoy
- **Paper :https://medium.com/@kashsala/building-samkash-tolstoy-a-tiny-lora-llm-that-lives-and-breathes-russian-literature-ca959747af4a
- **Demo:** 

## Attribution & Basics

- **Funded by:** Self-funded (individual project)
- **Shared by:** Samiya Kashif & Kashif Salahuddin (SamKash)
- **Model type:** LoRA (PEFT) adapter for a decoder-only causal language model (Qwen-family base, 1.5B params)
- **Language(s) (NLP):** English (`en`) — trained on English texts/metadata tagged as *Russian Literature* from Project Gutenberg
- **License:** `other` for the adapters (base model license applies separately)
- **Finetuned from model:** `deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B`

---

## Uses

### Direct Use
- Stylized long-form **generation** in the voice and conventions of 19th-century **Russian literature**.
- Brainstorming themes, motifs, and character interiority.
- Style-transfer scaffolding (draft → “make it sound like 19th-century Russian prose”).

### Downstream Use 
- As a component in a creative-writing assistant or editor plugin.
- Further **instruction tuning (SFT)** for tasks like summarization, theme extraction, or literature Q&A.
- Educational demos of domain-adaptive pretraining (DAPT) using LoRA/PEFT.

### Out-of-Scope Use
- Factual or safety-critical tasks (medical, legal, financial advice).
- Producing or implying authorship of genuine Tolstoy text.
- Modern colloquial dialogue or code generation (not optimized for these).

---

## Bias, Risks, and Limitations

- **Stylistic bias:** Strong tilt toward 19th-century Russian prose (long sentences, moral reflection).
- **Content bias:** Public-domain texts may reflect **outdated social views**.
- **Hallucination:** As a generative model, it can invent details; don’t use for factual claims.
- **Language scope:** Focused on English; performance on other languages is not guaranteed.

### Recommendations
- Keep a **human in the loop** for editing and intent verification.
- Avoid representing outputs as genuine text by historical authors.
- For classroom settings, clearly label generated content as synthetic.

---

## How to Get Started with the Model

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
from peft import PeftModel

base_id = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
adpt_id = "salakash/SamKash-Tolstoy"  # or local folder

device = "mps" if torch.backends.mps.is_available() else "cpu"
dtype  = torch.float16 if device == "mps" else torch.float32

tok = AutoTokenizer.from_pretrained(base_id, use_fast=True)
base = AutoModelForCausalLM.from_pretrained(base_id, dtype=dtype)
base.to(device)

model = PeftModel.from_pretrained(base, adpt_id)
model.config.use_cache = True  # inference

gen = pipeline("text-generation", model=model, tokenizer=tok, device=-1)
print(gen(
    "Write a reflective paragraph about conscience and fate in an aristocratic household.",
    max_new_tokens=200, do_sample=True, temperature=0.7, top_p=0.9
)[0]["generated_text"])