HMC83 commited on
Commit
90c6792
·
verified ·
1 Parent(s): cc85223

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +49 -7
README.md CHANGED
@@ -4,18 +4,60 @@ tags:
4
  - text-generation-inference
5
  - transformers
6
  - unsloth
7
- - llama
8
  license: apache-2.0
9
  language:
10
  - en
11
  ---
 
12
 
13
- # Uploaded finetuned model
 
14
 
15
- - **Developed by:** HMC83
16
- - **License:** apache-2.0
17
- - **Finetuned from model :** HuggingFaceTB/SmolLM2-360M-Instruct
18
 
19
- This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
 
 
 
 
 
 
 
20
 
21
- [<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  - text-generation-inference
5
  - transformers
6
  - unsloth
 
7
  license: apache-2.0
8
  language:
9
  - en
10
  ---
11
+ ## Model Description
12
 
13
+ Request Writer Smol has been fine tuned to generate Freedom of Information (FOI) requests to UK public authorities based on the autority name and three keywords. The model has been trained on a synthetic dataset of FOI requests covering various topics and public authorities across the UK.
14
+ The Model demonstrates improved generation of properly formatted, focused FOI requests for specific information that are unlikely to be refused on cost grounds.
15
 
16
+ ## Model Architecture
 
 
17
 
18
+ - **Base Model**: SmolLM2-360M-Instruct
19
+ - **Fine-tuning Method**: LoRA
20
+ - **LoRA Configuration**:
21
+ - Rank (r): 8
22
+ - Alpha: 16
23
+ - Dropout: 0.1
24
+ - Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
25
+ - **Training Parameters**: 2.34% of total parameters trained (8.68M trainable parameters)
26
 
27
+ ## Finetune training Data
28
+
29
+ ### Dataset Details
30
+ - **Source**: Synthetic FOI requests dataset (HMC83/synthetic_foi_requests)
31
+ - **Size**: 51,308 training examples, ~5,700 validation examples
32
+ - **Format**: Conversational format with system prompts, user inputs, and assistant responses
33
+
34
+ ### Training Configuration
35
+ - **Epochs**: 3
36
+ - **Batch Size**: 32
37
+ - **Learning Rate**: 1e-5
38
+ - **Optimizer**: AdamW 8-bit
39
+ - **Sequence Length**: 4096 tokens
40
+
41
+ ## Limitations and Considerations
42
+
43
+ Small size of the model (360M parameters) may limit the complexity of any generated requests. The model is trained specifically for UK FOI requests. It has not been trained to generate requests for information about individuals.
44
+
45
+ ## Usage Guidelines
46
+
47
+ ### Input Format
48
+ The model expects a prompt in the form of:
49
+ ```
50
+ Generate a formal Freedom of Information request to [authority_name] using these keywords: [keyword1, keyword2, keyword3]
51
+ ```
52
+
53
+ ### Output Format
54
+ It will try to generate a concinse, properly structured FOI request, starting with the phrase "Please provide me with a copy of the following information:" followed by 1 to 3 Numbered, specific information requests
55
+
56
+ ## Model Versions
57
+
58
+ ### Available Formats
59
+ - **LoRA Adapters**: `HMC83/request_writer_smol_lora`
60
+ - **Merged 16-bit**: `HMC83/request_writer_smol`
61
+
62
+ ### Disclaimer
63
+ Users are responsible for ensuring that their intended use complies with any applicable laws and regulations. Generated requests should be reviewed and potentially modified before submission to public authorities. Requests should be made in good faith and for legitimate purposes. The model can hallucinate, so any outputs should not be relied upon without being verified. Outputs may also reflect any biases that are present in the underlying training data.