Qwen3.5-9B-TLive

Livestream-domain SFT models at the 9B scale. Both variants see the same 10,000 training samples and the same recipe; they differ only in where fine-tuning starts — one from Qwen3.5-9B-Base, the other from the instruction-tuned Qwen3.5-9B. They live in a single repository, separated by subfolder.

Repository layout

v0_2/
├── Qwen3.5-9B-Base-SFT-10k    # fine-tuned from Qwen3.5-9B-Base
└── Qwen3.5-9B-chat-SFT-10k    # fine-tuned from Qwen3.5-9B (Instruct)

Loading

Pick a variant with subfolder:

from transformers import AutoModelForImageTextToText, AutoProcessor

repo = "L1ght/Qwen3.5-9B-TLive"
sub = "v0_2/Qwen3.5-9B-Base-SFT-10k"      # or v0_2/Qwen3.5-9B-chat-SFT-10k

model = AutoModelForImageTextToText.from_pretrained(
    repo, subfolder=sub, dtype="bfloat16", device_map="auto"
)
processor = AutoProcessor.from_pretrained(repo, subfolder=sub)

For text-only use, AutoModelForCausalLM + AutoTokenizer is lighter:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    repo, subfolder=sub, dtype="bfloat16", device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(repo, subfolder=sub)

messages = [{"role": "user", "content": "把这段直播内容总结一下：……"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)
out = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

vLLM cannot point at a subfolder directly; download one variant and pass the local path:

from huggingface_hub import snapshot_download
from vllm import LLM

path = snapshot_download(repo, allow_patterns=f"{sub}/*")
llm = LLM(model=f"{path}/{sub}")

Results

Comparison against the original Qwen3.5-9B-Base and Qwen3.5-9B (Instruct). Best per column in bold.

Livestream domain (text-only, 7 tasks)

Model	Reply Sel. R@1	Addressee	LiveQA	WhisperX Cont.	Live-CC Cont.	ShoppingMMLU	Sensai F1	Live_AVG
Qwen3.5-9B-Base	20.4	17.8	53.7	98.0	98.0	66.43	38.33	56.09
Qwen3.5-9B (Instruct)	48.2	47.6	60.5	99.8	99.4	69.04	47.25	67.40
Base-SFT-10k	62.2	59.8	56.2	100.0	100.0	70.81	42.99	70.29
chat-SFT-10k	63.6	59.0	55.8	100.0	99.8	71.62	45.40	70.75

General multimodal benchmarks

Model	VisuLogic	LongVideoBench	Video-MME	MVBench(19)	MMBench	MME*	General_AVG	Total_AVG
Qwen3.5-9B-Base	8.8	64.17	67.44	67.86	85.91	2362.2	58.84	57.47
Qwen3.5-9B (Instruct)	7.4	58.79	67.15	68.88	84.45	2394.8	57.33	62.37
Base-SFT-10k	9.1	62.90	68.22	70.70	87.37	2386.4	59.66	64.98
chat-SFT-10k	9.2	62.75	67.89	70.36	86.00	2406.8	59.24	65.00

Math reasoning

Model	AIME24	AIME25	GSM8K	MATH-500	Minerva	Math_AVG
Qwen3.5-9B-Base	20.00	20.00	88.63	68.20	37.87	46.94
Qwen3.5-9B (Instruct)	23.33	23.33	83.93	67.80	36.40	46.96
Base-SFT-10k	16.67	23.33	89.99	67.60	36.76	46.87
chat-SFT-10k	26.67	23.33	92.95	72.00	38.97	50.78

Zero-shot chain-of-thought, one greedy pass per question, scored with math_verify on the \boxed{} answer -- no external judge anywhere. AIME24 and AIME25 are 30 questions each, so one question moves the score by 3.33 points; read them as a coarse signal, not a ranking.

The learning rate that protects the general benchmarks protects maths too: the 1e-5 runs not released here fall to Math_AVG 36.96 (10k) and 31.92 (30k), with Minerva dropping from 37.87 to 20.59, while the 3e-6 Base-SFT released here stays level with the base model and the chat-start variant beats every 9B row measured.

Live_AVG is the mean of the 7 livestream-domain tasks; General_AVG is the mean of the 5 general benchmarks (VisuLogic, LongVideoBench, Video-MME, MVBench, MMBench); Total_AVG = (Live_AVG + General_AVG) / 2. *MME is the raw total score (max 2800) and is not included in the averages. Math_AVG is the mean of the 5 maths suites and is reported on its own: it is deliberately kept out of General_AVG and Total_AVG, so those two stay comparable with the numbers published before the maths suites were added.

Notes

SFT was text-only. Multimodal capability comes from the base model and received no additional training.

At this scale the learning rate matters more than it does at 2B/4B: the 1e-5 used for the smaller models costs several points on the general benchmarks, so both variants released here use 3e-6. The exact hyper-parameters of each run are in the train_config.json inside its subfolder.