Qwen2.5-7B-ChatCoder

A merged model combining Qwen2.5-7B-Instruct and Qwen2.5-Coder-7B-Instruct

---

Model Summary

Qwen2.5-7B-ChatCoder is a linearly merged language model that combines the instruction-following strength of Qwen2.5-7B-Instruct with the code generation capabilities of Qwen2.5-Coder-7B-Instruct.

The merge uses an 85% instruct / 15% coder weight split, carefully tuned to preserve the full chat and reasoning behaviour of the instruct model while absorbing coding knowledge from the coder model — resulting in a model that handles both natural conversation and code generation in a single set of weights.

Property	Value
Parameters	7.6B
Architecture	Qwen2ForCausalLM
Context length	128K tokens
Merge method	Linear
Instruct weight	0.85
Coder weight	0.15
dtype	bfloat16
Vocabulary size	152,064

---

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "ragunath-ravi/Qwen2.5-7B-ChatCoder"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

messages = [
    {"role": "system", "content": "You are a helpful coding assistant."},
    {"role": "user",   "content": "Write a Python function to do binary search."},
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=False,
        temperature=None,
        top_p=None,
        repetition_penalty=1.1,
        eos_token_id=[151645, 151643],
        pad_token_id=151645,
    )

response = tokenizer.decode(
    output[0][inputs.input_ids.shape[1]:],
    skip_special_tokens=True
)
print(response)

---

Run with 4-bit Quantization (5GB VRAM)

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

model_id = "ragunath-ravi/Qwen2.5-7B-ChatCoder"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto",
)

---

Hardware Requirements

Precision	VRAM Required	Recommended GPU
bfloat16 (full)	~16 GB	RTX 3090 / A100 / H100
8-bit (bitsandbytes)	~8 GB	RTX 3080 / 4080
4-bit NF4 (bitsandbytes)	~5 GB	RTX 3060 / 4060

---

What This Model Is Good At

• Instruction following — multi-turn chat, answering questions, reasoning
• Code generation — Python, JavaScript, C++, Java, SQL, Bash
• Code explanation — walk through what a piece of code does
• Debugging — find and fix bugs with explanation
• Mixed tasks — "explain this code and rewrite it to be more efficient"
• Math reasoning — step-by-step problem solving

---

Merge Details

This model was created using mergekit with a linear merge strategy.

models:
  - model: Qwen/Qwen2.5-7B-Instruct
    parameters:
      weight: 0.85
  - model: Qwen/Qwen2.5-Coder-7B-Instruct
    parameters:
      weight: 0.15

merge_method: linear
dtype: bfloat16

Why linear merge? Linear merging computes a direct weighted average of all model weights with no pruning or masking. It is the most stable merge method for combining an instruct-tuned model with a base/specialist model, avoiding the weight corruption that DARE-TIES can introduce when density < 1.0.

Why 85/15 split? The Qwen2.5-Coder-7B at this path is a base model (not instruct-tuned). At higher coder weights (tested at 0.4), the base model behaviour dominates and breaks instruction following. At 0.15, the coding knowledge is absorbed while the instruct fine-tuning remains intact.

---

Limitations

• Capabilities are bounded by the two parent models — this is not a trained model
• Very long code generation (>500 lines) may degrade in quality
• Not fine-tuned for agent/tool-use tasks
• May occasionally produce confident but incorrect code — always test generated code

---

Citation

If you use this model, please cite the original Qwen2.5 models:

@misc{qwen2.5,
  title  = {Qwen2.5: A Party of Foundation Models},
  author = {Qwen Team},
  year   = {2024},
  url    = {https://qwenlm.github.io/blog/qwen2.5/}
}

---

Created By

Merged and released by ragunath-ravi.

Built with mergekit by Arcee AI.