LFM2-2.6B-CodeAgent-LoRA

A LoRA fine-tuned adapter for LiquidAI/LFM2-2.6B-Exp trained to follow the smolagents CodeAgent format.

Model Description

This adapter teaches LFM2-2.6B-Exp to respond in the structured Thought + Code format required by smolagents CodeAgent:

Thought: I need to calculate this.
```python
result = 2 + 2
final_answer(result)
```

Key Features

• Base Model: LiquidAI/LFM2-2.6B-Exp (2.6B parameter hybrid architecture with LIV convolution + GQA)
• Format Compliance: 100% with minimal prompt
• Answer Accuracy: 80% on evaluation tasks
• Adapter Size: ~49MB (LoRA rank=8, alpha=16)

Training Details

Training Data

• 130 successful CodeAgent trajectories generated using Claude 3.5 Sonnet as the teacher model
• Tasks include mathematical reasoning, string manipulation, and general problem-solving
• Each trajectory demonstrates the Thought -> Code -> Observation -> final_answer pattern

Training Configuration

Parameter	Value
LoRA Rank	8
LoRA Alpha	16
Target Modules	q_proj, v_proj
Trainable Parameters	12.2M (0.47% of base)
Training Steps	30
Learning Rate	2e-4
Batch Size	4
Max Sequence Length	2048
Hardware	NVIDIA RTX 3090 (24GB)
Training Time	~5.5 hours

Training Framework

• TRL SFTTrainer
• PEFT for LoRA

Evaluation Results

Prompt Mode Comparison

Prompt Mode	Format Compliance	Answer Accuracy
Minimal	100%	80%
Default	80%	80%
None	0%	0%

The model performs best with the minimal prompt (~95 tokens), demonstrating successful prompt distillation.

Minimal Prompt Template

You are a CodeAgent that solves tasks by writing and executing Python code.

Always respond with Thought + Python code block. Example:

Thought: I need to calculate this.
```python
result = 2 + 2
final_answer(result)
```

Call final_answer(result) when done. Now Begin!

Usage

With PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "LiquidAI/LFM2-2.6B-Exp",
    device_map="auto",
    torch_dtype="bfloat16",
)
tokenizer = AutoTokenizer.from_pretrained("LiquidAI/LFM2-2.6B-Exp")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "krzysztofwos/LFM2-2.6B-CodeAgent-LoRA")

# Generate
messages = [{"role": "user", "content": "What is 15 * 23?"}]
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=0.3,
    min_p=0.15,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With smolagents

from smolagents import CodeAgent, FinalAnswerTool, TransformersModel

model = TransformersModel(
    model_id="LiquidAI/LFM2-2.6B-Exp",
    peft_model="krzysztofwos/LFM2-2.6B-CodeAgent-LoRA",
)

agent = CodeAgent(
    tools=[FinalAnswerTool()],
    model=model,
)

result = agent.run("What is 15 * 23?")
print(result)

Intended Use

• Code-assisted problem solving
• Mathematical reasoning tasks
• Automated code generation following structured formats
• Research into prompt distillation and small model fine-tuning

Limitations

• Requires specific prompt format: Works best with minimal prompt template
• Limited reasoning depth: 2.6B parameter hybrid architecture with LIV convolution + GQA model has constrained reasoning capabilities compared to larger models
• English only: Trained on English-language tasks

Citation

If you use this model, please cite:

@misc{lfm2_2.6b_codeagent_lora,
  author = {krzysztofwos},
  title = {LFM2-2.6B-CodeAgent-LoRA},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/krzysztofwos/LFM2-2.6B-CodeAgent-LoRA}
}

Acknowledgments

• LiquidAI for the LFM2-2.6B-Exp base model
• Hugging Face for smolagents, TRL, and PEFT
• Training performed as part of CodeAgent prompt distillation research