wesjos

Mimo-Qwen3.5-9B-Cold

README

License: apache-2.0

📋 Model Overview

Table with columns: Attribute, Value
Attribute	Value
Model Name	wesjos/Mimo-Qwen3.5-9B-Cold
Base	MiMo-V2.6-Distill-Qwen-9B (Qwen3.5-architecture distill, 9.4B params)
Method	DPO (Direct Preference Optimization), skipping SFT
Training Data	`lima_dpo_clean.jsonl` — 1,027 LIMA-style preference pairs
Training Framework	Unsloth + TRL DPOTrainer (QLoRA 4bit)
Release Formats	HF merged_16bit (18 GB) · LoRA adapter · GGUF Q8_0 (8.87 GB)
Context Length	262,144 (native)
Languages	English / Chinese

📊 Evaluation Results

Base Model Comparison (seed 42, temp 0, limit 200, evalscope 1.12)

Table with columns: Benchmark, Base, Cold, Δ
Benchmark	Base	Cold	Δ
ARC (Easy/Challenge)	91.25 (94.50/88.00)	94.75 (97.50/92.00)	+3.50 ↑
GSM8K	94.00	93.00	-1.00 (within noise)
MMLU	88.60	88.16	-0.44 (within noise)
BBH	82.41	79.63	-2.78 (high sampling noise at limit)

Code & Tool Calling (500 ToolBench-Static: in_domain 333 + out_of_domain 167)

Table with columns: Metric, Base, Cold, Δ
Metric	Base	Cold	Δ
HumanEval pass@1 (full 164)	65.85	73.78	+7.93 ↑
ToolBench Plan.EM	77.75	77.54	-0.21 (flat)
ToolBench Act.EM	17.71	17.93	+0.22 (flat)
ToolBench F1	15.16	15.17	±0 (flat)

Conclusion

✅ Style imprint confirmed: concise and decisive ✅ Intelligence preserved: knowledge/reasoning/instruction-following flat or improved ✅ Code ability notably improved: HumanEval +7.9pp ✅ Tool planning retained: Plan.EM/F1 flat ⚠️ Side effect: tool hallucination rate +4.5pp (decisive style makes the model hesitate less when uncertain) — agent use cases should add schema validation as a fallback

🚀 Usage

Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained(    "<repo>/merged_16bit", torch_dtype="bfloat16", device_map="auto")tok = AutoTokenizer.from_pretrained("<repo>/merged_16bit") messages = [{"role": "user", "content": "Introduce yourself in two sentences."}]inputs = tok.apply_chat_template(    messages, tokenize=True, add_generation_prompt=True,    enable_thinking=False, return_tensors="pt", return_dict=True,).to("cuda")out = model.generate(**inputs, max_new_tokens=300, temperature=0.6, top_p=0.9)print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

llama.cpp (GGUF Q8_0)

bash

llama-server -m mimo-9b-cold-Q8_0.gguf \  --ctx-size 8192 -ngl 99 -ctk q8_0 -ctv q8_0 -fa on \  --jinja --alias mimo-9b-cold --port 18080

Recommended Parameters

temperature: 0.6, top_p: 0.9, top_k: 20
Anti-repetition: repeat_penalty 1.05, DRY multiplier 0.8
Thinking mode: template supports enable_thinking toggle

⚠️ Limitations & Notes

Small training scale: only 1 epoch × 1,027 pairs — lightweight style alignment; do not expect capability gains, only style transfer
Slightly higher tool hallucination rate: for function calling, add parameter schema validation at the application layer
Knowledge cutoff: inherited from the base model (Qwen3.5 family); training data contains no new knowledge
Safety alignment: retains the base model's full safety values (illegal requests are refused with lawful alternatives)
BBH sampling variance: at limit 200 each subset has only ~4 questions; the -2.78pp drop is not statistically significant

📁 File Listing

text

├── merged_16bit/            # HF full weights (bfloat16, 4 shards)│   ├── model-0000X-of-00004.safetensors│   ├── config.json / generation_config.json│   ├── tokenizer.json / tokenizer_config.json│   └── chat_template.jinja├── adapter/                 # LoRA adapter (r=16, 0.04% params)└── mimo-9b-cold-Q8_0.gguf # llama.cpp Q8_0 quant (8.87 GB)

🙏 Acknowledgements

Base: MiMo-V2.6-Distill-Qwen-9B (Qwen3.5 architecture)
Methods: LIMA (Less Is More for Alignment) · DPO (Rafailov et al.)
Tools: Unsloth · TRL · evalscope · llama.cpp

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

wesjos

Model Tree

Base

XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality