---

base_model:

• Qwen/Qwen3-0.6B
• prithivMLmods/rStar-Coder-Qwen3-0.6B
• suayptalha/Qwen3-0.6B-IF-Expert

datasets:

• microsoft/rStar-Coder
• patrickfleith/instruction-freak-reasoning

Qwen3-0.6B-rStar-Coder-IF-Expert

A high-skill merged model created with mergekit that combines rStar-Coder's expert code generation with IF-Expert's precise instruction following via SLERP interpolation.

Merge Methodology

Property	Value
Merge Technique	SLERP (Spherical Linear Interpolation)
Base Models	prithivmlmods/rStar-Coder-Qwen3-0.6B(coding specialization)<br>•suayptalha/Qwen3-0.6B-IF-Expert` (instruction fidelity)
Architecture	Qwen3 0.6B decoder-only transformer
Layer Strategy	Full 28-layer SLERP merge with parameter-specific weights

Merge Configuration

base_model: prithivmlmods/rStar-Coder-Qwen3-0.6B
dtype: float16
merge_method: slerp
parameters:
  t:
    - filter: embed_tokens
      value: 0.0          # 100% rStar-Coder embeddings
    - filter: self_attn
      value: 0.5          # 50/50 attention interpolation
    - filter: mlp
      value: 0.5          # 50/50 MLP interpolation
    - filter: lm_head
      value: 1.0          # 100% IF-Expert output head
    - value: 0.5          # Default 50/50 for remaining params
slices:
  - sources:
      - layer_range: [0, 28]
        model: prithivmlmods/rStar-Coder-Qwen3-0.6B
      - layer_range: [0, 28]
        model: suayptalha/Qwen3-0.6B-IF-Expert

Capabilities

This merge strategically balances:

• ✅ rStar-Coder DNA: Production-grade Python/TS code generation with algorithmic precision
• ✅ IF-Expert DNA: Strict adherence to complex instructions and format constraints
• ✅ Emergent strength: Better instruction-compliant code than either base model alone

Ideal Use Cases

# Example: Complex multi-constraint request
query = "Write a FastAPI endpoint that:
1. Accepts CSV uploads
2. Validates rows against Pydantic schema
3. Returns paginated JSON
4. Includes OpenAPI docs
5. Uses async file handling"

# This model reliably satisfies all 5 constraints in one generation

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "gss1147/Qwen3-0.6B-rStar-Coder-IF-Expert"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)

messages = [{"role": "user", "content": "Write a thread-safe connection pool in Python"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.2,
    do_sample=True
)
code = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(code)

Important Notes

⚠️ Local Merge Origin
This model was merged locally from private checkpoints (X:/AI_Models/...). The weights are redistributed here under Apache 2.0 per base model licenses. No additional training data was used — this is purely a parameter-space merge.

License

Apache 2.0 (inherited from base Qwen3 models). No additional restrictions from merge process.

Citation

@software{gss1147_qwen3_merge_2026,
  author = {gss1147},
  title = {Qwen3-0.6B-rStar-Coder-IF-Expert},
  year = {2026},
  note = {SLERP merge of rStar-Coder and IF-Expert variants via mergekit},
  url = {https://huggingface.co/gss1147/Qwen3-0.6B-rStar-Coder-IF-Expert}
}