SmolLM2-135M-Eagle

SmolLM2-135M-Eagle is a fine-tuned version of the SmolLM2-135M model on the EagleSFT dataset, designed to improve the model's capabilities in both Russian and English language tasks.

GGUF version of this model is available at: SmolLM2-135M-Eagle-GGUF

Model Description

SmolLM2-135M-Eagle is a lightweight language model that has been fine-tuned specifically to handle bilingual content. This fine-tuning extends the base model's capabilities to better understand and generate content in Russian while maintaining its English competency.

Base Model

The model is built upon SmolLM2-135M, a compact language model with 135 million parameters that offers a good balance between performance and resource requirements.

Fine-tuning Details

Dataset

The model was fine-tuned on the EagleSFT dataset, which contains 536,231 pairs of human questions and machine-generated responses in both Russian and English languages. The dataset primarily focuses on educational content but also includes everyday questions and casual conversations.

Environmental Impact

• Training duration: 26h 26m total
  • 15h 19m 52s in Tyumen, Russia (300W power consumption)
  • 11h 6m 8s in Saint-Petersburg (360W power consumption)
• Hardware: 1 x RTX 4090
• Carbon emissions: Approximately 3.11 kg CO2eq
  • Calculated based on average power consumption and average CO2eq/kWh (350g) in these regions
  • Tyumen: 300W * 15.33h * 350g/kWh = 1.61 kg CO2eq
  • Saint-Petersburg: 360W * 11.10h * 350g/kWh = 1.50 kg CO2eq

Training Parameters

• Training approach: Supervised Fine-Tuning (SFT)
• Training epochs: 2
• Learning rate: 3.0e-04
• Precision: bfloat16

Limitations and Capabilities

It's important to note that this model was not pre-trained but only underwent SFT on a relatively small number of tokens. This means that the model has a limited amount of data to rely on when answering in Russian compared to its English capabilities.

Despite extensive limitations, the model shows minimal improvement in:

• Basic recognition of Russian prompts (though with frequent misunderstandings)
• Handling simple tasks formatted as "{question in Russian}, answer in English"
• Basic translation from Russian to English (though quality remains poor)

The model's minimal understanding of Russian language comes solely from the supervised fine-tuning process without any proper pre-training with Russian text corpus, resulting in severely limited capabilities.

Experimental Capabilities

The model demonstrates some experimental capabilities, but with significant limitations:

• Basic Russian text understanding (with frequent errors and misinterpretations)
• Limited question answering in Russian (quality significantly lower than English)
• Basic Russian to English translation (better than English to Russian)

Limitations

• NOT SUITABLE FOR PRODUCTION USE: This model should not be used in production environments in any form
• Extremely limited knowledge base for Russian language due to lack of pre-training with Russian text
• Unoptimized tokenizer performance for Russian language results in inefficient token usage
• Output quality in Russian will be unsatisfactory for most use cases
• May produce inaccurate, inconsistent, or inappropriate responses, especially in Russian
• All limitations of the base SmolLM2-135M model still apply