Instructions to use marianna13/prelude-9b-sft-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use marianna13/prelude-9b-sft-v2 with Transformers:

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

pipe = pipeline("text-generation", model="marianna13/prelude-9b-sft-v2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("marianna13/prelude-9b-sft-v2")
model = AutoModelForCausalLM.from_pretrained("marianna13/prelude-9b-sft-v2", 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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use marianna13/prelude-9b-sft-v2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "marianna13/prelude-9b-sft-v2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "marianna13/prelude-9b-sft-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/marianna13/prelude-9b-sft-v2

SGLang

How to use marianna13/prelude-9b-sft-v2 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 "marianna13/prelude-9b-sft-v2" \
    --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": "marianna13/prelude-9b-sft-v2",
		"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 "marianna13/prelude-9b-sft-v2" \
        --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": "marianna13/prelude-9b-sft-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use marianna13/prelude-9b-sft-v2 with Docker Model Runner:
```
docker model run hf.co/marianna13/prelude-9b-sft-v2
```

Prelude 9B SFT v2 — `step9020`

Supervised fine-tune of marianna13/prelude-9b-midtrain-v1 (revision stage4-128k-step1001: 128K context, YaRN x2 over 64K, RoPE theta 9M). Qwen3 architecture, vocab 262,272, Prelude ChatML chat template (Qwen3-compatible: <think> reasoning, hermes <tool_call> tool calls; enable_thinking=False prefills an empty think block).

Final checkpoint: Megatron iter 9020, one pass over the 37.7B-token SFT mix (intermediate exports are not published).

Data (37.69B tokens, 4.34M conversations, one pass, packed at 64K)

Source	Rows	Tokens
INTELLECT-3-SFT math (`openreasoning_math`, <= 4 traces per prompt)	634,354	11.97B
INTELLECT-3-SFT code (<= 12 per prompt)	406,380	6.91B
INTELLECT-3-SFT science (53% sample)	847,922	5.48B
INTELLECT-3-SFT chat + IF (`am_chat`, `am_if`)	1,005,318	2.37B
INTELLECT-3-SFT native tool calls (`openreasoning_tool`, `toucan_tool`)	422,438	1.59B
INTELLECT-3-SFT `swe_swiss`	9,934	0.17B
open-thoughts/AgentTrove, Terminus-2 agent episodes (behavioural filter, teacher-balanced)	595,509	7.29B (2.48B trained)
nebius/SWE-rebench-openhands-trajectories (native-tool SWE agent)	20,256	1.04B (0.26B trained)
Thinking-off IF/chat (reasoning stripped, empty think block)	398,253	0.89B

Rows overlapping the eval sets and RL task pools were removed (decontamination by exact prompt keys).

Training

Megatron-Bridge on MareNostrum 5, 64 nodes, global batch 64 x 64K tokens, LR 2e-5 cosine, 9,020 steps (one pass).

Evaluation (`main` = step 9020)

Benchmark	think on	think off
AIME24 / AIME25	54.3 / 43.3	50.3 / 37.0
MATH500	84.6	78.2
GPQA-Diamond	51.2	45.6
IFEval (prompt-level loose)	77.3	74.9
HumanEval+	80.8	65.6
MultiPL-E (10 languages)	46.9	46.2
SWE-bench Verified (random 100), terminus-2	17.3
Terminal-Bench 2.1, terminus-2	3.0
TBLite (100), terminus-2	7.7

Evalchemy with a 32,768-token budget; thinking on sampled at T=0.6 / top-p 0.95, thinking off greedy. Agentic: mean of 3 runs, thinking on, 3x timeouts. For the full comparison (Qwen3.5-9B-Base, Qwen3.5-9B, data-scaling steps) see the rl/evals/prelude_midtrain/sft_v2_full_vs_qwen35.md table in the project monorepo.

Data scaling (intermediate exports of the same run, evaluated but not published):

Checkpoint	AIME24 think	HumanEval+ think	IFEval think	SWE / TB2.1 / TBLite
step 500 (2.1B tok)	36.3	68.3	71.0	14.0 / 1.9 / 6.0
step 1000 (4.2B)	41.0	70.1	70.0	12.3 / 1.9 / 6.7
step 2500 (10.5B)	46.0	72.0	73.4	14.7 / 1.9 / 7.7
step 5000 (21.0B)	48.0	78.0	75.8	16.0 / 3.4 / 8.3
step 9020 (37.7B, this model)	54.3	80.8	77.3	17.3 / 3.0 / 7.7

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("marianna13/prelude-9b-sft-v2")
m = AutoModelForCausalLM.from_pretrained("marianna13/prelude-9b-sft-v2", torch_dtype="bfloat16")
ids = tok.apply_chat_template([{"role": "user", "content": "Hi"}], add_generation_prompt=True,
                              enable_thinking=True, return_tensors="pt")

Recommended sampling with thinking on: temperature 0.6, top-p 0.95 (greedy decoding often fails to close the reasoning). Research checkpoint; dataset licences follow the upstream sources listed above.

Downloads last month: 176

Safetensors

Model size

9B params

Tensor type

BF16

Model tree for marianna13/prelude-9b-sft-v2

Base model

marianna13/prelude-9b-midtrain-v1

Finetuned

(1)

this model