---
library_name: peft
tags:
- tiiuae
- code
- instruct
- databricks-dolly-15k
- falcon-40b
datasets:
- databricks/databricks-dolly-15k
base_model: tiiuae/falcon-40b
license: apache-2.0
---

### Finetuning Overview:

**Model Used:** tiiuae/falcon-40b  
**Dataset:** Databricks-dolly-15k  

#### Dataset Insights:
The Databricks-dolly-15k dataset, comprising over 15,000 records, stands as a testament to the dedication of numerous Databricks professionals. Aimed at refining the interactive capabilities of systems like ChatGPT, the dataset offers:

- Prompt/response pairs across eight distinct instruction categories.
- A blend of the seven categories from the InstructGPT paper and an open-ended category.
- Original content, devoid of generative AI influence and primarily offline-sourced, with exceptions for Wikipedia references.
- Interactive sessions where contributors could address and rephrase peer questions.

Note: Some data categories incorporate Wikipedia references, evident from bracketed citation numbers, e.g., [42]. Exclusion is recommended for downstream applications.

#### Finetuning Details:

Leveraging [MonsterAPI](https://monsterapi.ai)'s no-code [LLM finetuner](https://docs.monsterapi.ai/fine-tune-a-large-language-model-llm), our finetuning emphasized:

- **Cost-Effectiveness:** A complete run at just `$11.8`.
- **Efficiency:** Using an A6000 48GB GPU, the session concluded in 5 hours and 40 minutes.

#### Hyperparameters & Additional Details:

- **Epochs:** 1
- **Learning Rate:** 0.0002
- **Data Split:** Training 90% / Validation 10%
- **Gradient Accumulation Steps:** 4

---

### Prompt Structure:
```
### INSTRUCTION:
[instruction]

[context]

### RESPONSE:
[response]
```

Loss metrics

Training loss:
![training loss](train-loss.png "Training loss")


---

license: apache-2.0