Text Generation
Transformers
Safetensors
Turkish
qwen3
turkish
causal-lm
from-scratch
pretraining
text-generation-inference
Instructions to use eademir/TurkLM-scratch-0.3B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use eademir/TurkLM-scratch-0.3B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="eademir/TurkLM-scratch-0.3B")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("eademir/TurkLM-scratch-0.3B") model = AutoModelForCausalLM.from_pretrained("eademir/TurkLM-scratch-0.3B", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use eademir/TurkLM-scratch-0.3B with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "eademir/TurkLM-scratch-0.3B" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "eademir/TurkLM-scratch-0.3B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/eademir/TurkLM-scratch-0.3B
- SGLang
How to use eademir/TurkLM-scratch-0.3B with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "eademir/TurkLM-scratch-0.3B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "eademir/TurkLM-scratch-0.3B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "eademir/TurkLM-scratch-0.3B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "eademir/TurkLM-scratch-0.3B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use eademir/TurkLM-scratch-0.3B with Docker Model Runner:
docker model run hf.co/eademir/TurkLM-scratch-0.3B
File size: 3,394 Bytes
f354976 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 | """
Bits-per-character (BPC) of a causal LM on two sets of Turkish Wikipedia articles.
overlap : Wikipedia articles #5,001-5,200 of the 20231101 dump (data/bpc_docs.json, cached).
WARNING: this slice lies inside the TurkLM training stream - it is NOT held out. It is kept only because the
first published numbers used it; it rewards models that saw it. Do not use it to compare TurkLM with other models.
fresh : articles created AFTER the training cutoff (data/fresh_wiki.json, built by fetch_fresh_wiki.py). Use this one.
BPC = total negative log-likelihood (bits) / number of characters. It is comparable across tokenizers (perplexity is not).
Method: each document is cut into windows of 1,024 tokens; a window needs >= 2 tokens; loss is the mean token NLL of the window.
IMPORTANT (macOS / Apple GPU): the default sdpa attention on MPS (macOS 27, torch 2.12.1) is not causal and gives wrong,
far too low BPC. This script always uses attn_implementation="eager", which matches CPU fp32 and CUDA.
usage: python bpc_two_sets.py --model-path MODEL_DIR_OR_HF_ID [--json out.json] [--device cpu|cuda|mps]
"""
import argparse, json, math
from pathlib import Path
import torch
from tqdm import tqdm
from transformers import AutoModelForCausalLM, AutoTokenizer
HERE = Path(__file__).parent
MAX_SEQ_LEN = 1024
@torch.no_grad()
def doc_nats_chars(model, tok, text, device):
text = text.strip()
ids = tok.encode(text, add_special_tokens=False)
nats = 0.0
for i in range(0, len(ids), MAX_SEQ_LEN):
chunk = ids[i:i + MAX_SEQ_LEN]
if len(chunk) < 2:
continue
x = torch.tensor([chunk], device=device)
nats += model(input_ids=x, labels=x).loss.item() * (len(chunk) - 1)
return nats, len(text)
def bpc(model, tok, texts, device, desc):
per_doc = [doc_nats_chars(model, tok, t, device) for t in tqdm(texts, desc=desc)]
return sum(n for n, _ in per_doc) / (math.log(2) * sum(c for _, c in per_doc)), per_doc
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--model-path", required=True)
ap.add_argument("--json", default="")
ap.add_argument("--device", default="")
ap.add_argument("--sets", default="overlap,fresh")
a = ap.parse_args()
device = torch.device(a.device or ("cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"))
tok = AutoTokenizer.from_pretrained(a.model_path)
model = AutoModelForCausalLM.from_pretrained(a.model_path, dtype=torch.bfloat16, attn_implementation="eager").to(device).eval()
out = {"model": a.model_path}
if "overlap" in a.sets:
texts = json.loads((HERE / "data" / "bpc_docs.json").read_text())
out["overlap"], per = bpc(model, tok, texts, device, "overlap set")
out["overlap_n"] = len(texts)
print(f"BPC overlap set ({len(texts)} articles, inside the training stream): {out['overlap']:.4f}")
if "fresh" in a.sets:
docs = json.loads((HERE / "data" / "fresh_wiki.json").read_text())
out["fresh"], per = bpc(model, tok, [d["text"] for d in docs], device, "fresh set")
out["fresh_n"] = len(docs)
print(f"BPC fresh set ({len(docs)} articles created after the cutoff): {out['fresh']:.4f}")
if a.json:
Path(a.json).write_text(json.dumps(out, indent=1))
if __name__ == "__main__":
main()
|