Instructions to use JayZenith/glyph-sft-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use JayZenith/glyph-sft-v1 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="JayZenith/glyph-sft-v1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("JayZenith/glyph-sft-v1")
model = AutoModelForCausalLM.from_pretrained("JayZenith/glyph-sft-v1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use JayZenith/glyph-sft-v1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "JayZenith/glyph-sft-v1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "JayZenith/glyph-sft-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/JayZenith/glyph-sft-v1

SGLang

How to use JayZenith/glyph-sft-v1 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 "JayZenith/glyph-sft-v1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "JayZenith/glyph-sft-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "JayZenith/glyph-sft-v1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "JayZenith/glyph-sft-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use JayZenith/glyph-sft-v1 with Docker Model Runner:
```
docker model run hf.co/JayZenith/glyph-sft-v1
```

glyph-sft-v1

SFT of Qwen/Qwen3-4B-Base on a custom TASK trace format (plans, todos, tool calls, satisfaction markers ⊨, response blocks). LoRA on attention + MLP, with lm_head in modules_to_save and a separate higher learning rate so termination tokens are actually learned.

Training data: JayZenith/glyph-sft-v1-data (private, 1098 traces, 80/10/10 split).

Held-out test loss (110 traces, never seen in training)

	base	sft	delta
mean loss	1.280	0.972	−0.308
perplexity	3.60	2.64	36% lower
SFT wins per ex.			110/110

Format quality (5-prompt generation eval)

	base	sft
valid trace	0/5	4/5
ends with response	0%	100%
has plan	0%	100%
no repetition	60%	100%
not truncated	20%	100%
used tools when given	0/4	4/4
avg score (out of 7)	0.2	6.4

The one failed valid-trace was a no-tool reasoning prompt where the model wrote a 5-step plan but didn't emit ⊨ N satisfaction markers for every step. This is a fixable issue and a clean target for RL.

Training

1× A100 80GB SXM4, ~1h32m
LoRA rank 64, alpha 64, dropout 0.05
targets: q,k,v,o,gate,up,down
modules_to_save=["lm_head"]
LR = 2e-5 (both trunk and lm_head — separate optimizer groups, equal rate)
assistant-only loss masking
3 epochs, 330 total steps, eval every 50 steps incl. greedy gen-eval
trainable: 521M / 4.54B params (11.5%)
final eval_loss: 0.958

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("JayZenith/glyph-sft-v1")
model = AutoModelForCausalLM.from_pretrained(
    "JayZenith/glyph-sft-v1", torch_dtype=torch.bfloat16, device_map="auto"
)

Status

This is the SFT starting point for an RL run (validator-shaped reward, prime-rl). It is not a final chat model.

Downloads last month: 901

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for JayZenith/glyph-sft-v1

Base model

Qwen/Qwen3-4B-Base

Adapter

(53)

this model