LFM25-1.2B-CodeAgent-opus4-default

A LoRA fine-tuned adapter for LiquidAI/LFM2.5-1.2B-Instruct trained to follow the smolagents CodeAgent format.

Model Description

This adapter teaches LFM2.5-1.2B 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/LFM2.5-1.2B-Instruct (1.17B parameters)
• Token Accuracy: 90.2% on training data
• Single-Turn Rate: 53.8% of tasks solved in one turn
• Adapter Size: ~42MB (LoRA rank=16, alpha=32)

Training Details

Training Data

• 130 successful CodeAgent trajectories generated using Claude Opus 4.5
• Tasks include mathematical reasoning, string manipulation, file operations, and general problem-solving
• Each trajectory demonstrates the Thought → Code → Observation → final_answer pattern
• Average trajectory length: 2971 tokens

Training Configuration

Parameter	Value
LoRA Rank	16
LoRA Alpha	32
Target Modules	w1, w2, w3, q_proj, k_proj, v_proj, out_proj, in_proj
Trainable Parameters	11.1M (0.94% of base model)
Epochs	3
Learning Rate	1e-4
Batch Size	2 (effective 8 with gradient accumulation)
Max Sequence Length	8192
Hardware	NVIDIA RTX 3090 (24GB)
Training Time	260s

Training Framework

• TRL SFTTrainer
• PEFT for LoRA

Ablation Study Results

This model is part of a teacher ablation study comparing different Claude models and prompting strategies:

Teacher Config	Token Accuracy	Training Time	Avg Trajectory Tokens
haiku-default	73.8%	461s	2,957
sonnet-default	90.6%	260s	3,054
opus4-default	90.2%	260s	2,971
sonnet4-terse	94.9%	93s	632
opus4-terse	95.0%	76s	613

Key Finding: Terse, focused trajectories from capable teachers transfer significantly better to small models.

Usage

With PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "LiquidAI/LFM2.5-1.2B-Instruct",
    device_map="auto",
    torch_dtype="bfloat16",
)
tokenizer = AutoTokenizer.from_pretrained("LiquidAI/LFM2.5-1.2B-Instruct")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "krzysztofwos/LFM25-1.2B-CodeAgent-opus4-default")

# 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=0.1,
    top_p=0.1,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With smolagents

from smolagents import CodeAgent, FinalAnswerTool, TransformersModel

model = TransformersModel(
    model_id="LiquidAI/LFM2.5-1.2B-Instruct",
    peft_model="krzysztofwos/LFM25-1.2B-CodeAgent-opus4-default",
)

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

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

Intended Use

• Code-assisted problem solving with small, efficient models
• Mathematical reasoning tasks
• Automated code generation following structured formats
• Research into prompt distillation and teacher model selection

Limitations

• 1.2B model constraints: Limited reasoning depth compared to larger models
• English only: Trained on English-language tasks
• CodeAgent format specific: Optimized for smolagents Thought/Code/final_answer pattern

Citation

If you use this model, please cite:

@misc{lfm25-codeagent-opus4_default-2025,
  author = {krzysztofwos},
  title = {LFM25-1.2B-CodeAgent-opus4-default},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/krzysztofwos/LFM25-1.2B-CodeAgent-opus4-default}
}

Acknowledgments

• LiquidAI for the LFM2.5 base model
• Anthropic for Claude Opus 4.5 (teacher model)
• Hugging Face for smolagents, TRL, and PEFT