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

Libraries

How to use Corning/qwen-sft with Transformers:

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

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

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Corning/qwen-sft")
model = AutoModelForCausalLM.from_pretrained("Corning/qwen-sft", device_map="auto")
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 Settings

vLLM

How to use Corning/qwen-sft with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/Corning/qwen-sft

SGLang

How to use Corning/qwen-sft 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 "Corning/qwen-sft" \
    --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": "Corning/qwen-sft",
		"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 "Corning/qwen-sft" \
        --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": "Corning/qwen-sft",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Corning/qwen-sft with Docker Model Runner:
```
docker model run hf.co/Corning/qwen-sft
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

qwen-sft

Full-parameter SFT of Qwen/Qwen3-32B on a 6,962-trajectory subset of DR-Tulu-style deep-research agent data (Datasearch/DR-AntiForget · dr_tulu_full_sft_6962.jsonl, link).

The model is intended as the SFT baseline for the DR-AntiForget project, which studies catastrophic forgetting on deep-research agent capabilities.

Training


Base	`Qwen/Qwen3-32B`
Method	full-parameter SFT
Framework	LLaMA-Factory + DeepSpeed ZeRO-3 (no offload) + FlashAttention-2 + Liger Kernel + sequence packing
Hardware	1 node × 8 × NVIDIA H200 (140 GB)
Wall time	19h 20m
Epochs	5
Optimizer	AdamW, lr 4e-5 (cosine, 10% warmup, weight decay 0)
Effective batch	64 (mb 1 × grad-accum 8 × 8 GPUs)
Cutoff length	32,768 tokens
Precision	bfloat16
Trainable params	32.76 B
Total steps	285 (with packing; 1.92× compaction vs. raw)

Data preprocessing

Raw dr_tulu_full_sft_6962.jsonl mean length is ≈19k Qwen tokens; ~13% of trajectories exceed 32k. To preserve as much of each trajectory as possible under a 32k cutoff, long trajectories are pre-compressed in place: middle <tool_response> bodies are replaced with a small JSON placeholder ({omitted, original_tokens, query, preview}) while keeping the first 1 and last 5 tool responses intact. This brings the >32k tail from 13.3% to 1.8% of the dataset.

Loss

epoch	loss
0.09	0.6409
1.05	0.3295
2.99	0.1547
3.07	0.0855 (post-epoch-3 boundary drop)
4.04	0.0474
5.00	0.0257

Train loss only — no holdout was wired into this run, so be aware of likely overfit at epoch 5. If you have a held-out eval set, use it before deploying.

Quickstart

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

repo = "Corning/qwen-sft"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype=torch.bfloat16, device_map="auto",
    attn_implementation="flash_attention_2", trust_remote_code=True,
)

msgs = [
    {"role": "system", "content": "You are a helpful research assistant."},
    {"role": "user",   "content": "Summarize catastrophic forgetting in one paragraph."},
]
text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
ids  = tok(text, return_tensors="pt").to(model.device)
out  = model.generate(**ids, max_new_tokens=512, temperature=0.6, top_p=0.95, top_k=20)
print(tok.decode(out[0, ids.input_ids.shape[1]:], skip_special_tokens=False))

Tool-use format

The model was trained on Qwen-native tool-use:

Tool calls (assistant) wrapped in <tool_call>...JSON...</tool_call>
Tool results (sent back as a user-role message) wrapped in <tool_response>...JSON...</tool_response>
The tool_response span is masked out of the loss (the model only learns to use tool results, not to predict them)

The shipped chat_template.jinja renders messages with tools=[...] correctly.

Files

14 sharded safetensors (~64 GB bf16)
chat_template.jinja, full tokenizer files
config.json, generation_config.json
training_loss.png, train_results.json, all_results.json (training metadata)
inference.py (transformers + vLLM demo)

Framework versions

Transformers 4.52.4
Pytorch 2.6.0+cu124
Datasets 3.6.0
Tokenizers 0.21.1

License

Apache-2.0, inheriting from the base model.

Downloads last month: 2

Safetensors

Model size

33B params

Tensor type

BF16

Model tree for Corning/qwen-sft

Base model

Qwen/Qwen3-32B

Finetuned

(542)

this model