lugha1

lugha1

Model Summary

lugha1 is an instruction-tuned and preference-aligned conversational language model derived from SmolLM-1.7B-Instruct. It is optimized for clear instruction following, polite dialogue, and socially grounded responses, with an emphasis on small-talk, explanations, and helpful conversational tone.

The model was trained using Supervised Fine-Tuning (SFT) and further aligned using Direct Preference Optimization (DPO) with parameter-efficient fine-tuning (LoRA).

---

Model Details

• Base model: HuggingFaceTB/SmolLM-1.7B-Instruct

• Architecture: Decoder-only Transformer

• Parameters: ~1.7B (LoRA-adapted)

• Fine-tuning method:

  • QLoRA (4-bit quantization)
  • LoRA (PEFT)
  • SFT → DPO alignment

• Precision: 4-bit (training), fp16 compute

• Language(s): English (primary); conversational multilingual robustness inherited from base model

• Model name: lugha1

---

Training Procedure

Instruction Tuning (SFT)

The model was first trained using supervised instruction–response pairs, formatted in a conversational style:

### Instruction:
<user instruction>

### Response:
<assistant response>

This stage teaches the model:

• instruction comprehension
• response structure
• conversational coherence

Preference Alignment (DPO)

After SFT, the model was aligned using Direct Preference Optimization (DPO). Preference pairs (chosen vs. rejected responses) were used to bias the model toward:

• more helpful answers
• clearer explanations
• socially appropriate tone

DPO was chosen over RLHF to ensure:

• training stability
• reproducibility
• low computational overhead

---

Datasets

Training data consisted of instruction-style conversational datasets inspired by Smalltalk-style interactions, including:

• General instruction-following
• Explanatory responses
• Conversational and polite dialogue patterns

Note: Some preference data was synthetically generated for alignment purposes.

---

Intended Use

lugha1 is suitable for:

• Instruction-following chatbots
• Educational assistants
• Conversational agents
• Prototyping aligned LLM systems
• Research on small, efficient instruction-tuned models

Example use cases:

• Question answering
• Explanations
• Friendly dialogue
• Writing assistance

---

Out-of-Scope Use

This model is not intended for:

• Medical or legal advice
• High-stakes decision-making
• Fully autonomous agents without human oversight
• Safety-critical systems

---

Bias, Risks, and Limitations

• The model may reflect biases present in the training data.
• Preference alignment improves tone and helpfulness but does not guarantee factual correctness.
• Like all language models, it may hallucinate when uncertain.

Users should apply appropriate validation and human oversight.

---

Ethical Considerations

• No personal data was intentionally included in training.
• The model was trained using publicly available or synthetic instruction data.
• Alignment techniques were used to encourage polite and cooperative behavior.

---

How to Use

from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="your-username/lugha1",
    tokenizer="your-username/lugha1",
    device_map="auto"
)

prompt = """### Instruction:
Explain empathy using humor.

### Response:"""

pipe(prompt, max_new_tokens=150)

---

Training Infrastructure

• Platform: Google Colab (free tier)

• GPU: NVIDIA T4 / L4

• Frameworks:

  • 🤗 Transformers
  • 🤗 Datasets
  • TRL
  • PEFT
  • bitsandbytes

---

License

This model inherits the license of its base model: SmolLM-1.7B-Instruct license (see base model card for details).

---