Arnab-Saha-241562423/credit-card-usage-guide-lm
Tiny GPT-style model trained for educational credit-card usage guidance.
Model Details
- Architecture: decoder-only transformer (nanoGPT-style)
- Tokenization: character-level tokenizer (
char_tokenizer.json)
- Domain: retail banking education (credit-card usage)
Intended Use
- Educational Q&A about credit card usage basics.
- Not for personalized financial advice.
Evaluation Method
- Split: lightweight internal prompts from domain-focused examples.
- Evaluation style: manual and semi-structured checks.
- Metrics tracked:
- Perplexity from held-out text chunks.
- Key-point coverage on Q&A prompts.
- Safety behavior checks on risky prompts.
Benchmark Snapshot
| Metric | Method | Result |
|---|
| Perplexity (lower is better) | python main.py --mode eval --checkpoint checkpoints/sft_model.pt | TBD on user machine |
| Q&A Relevance | Manual review on sample prompts | Mostly relevant for common card-usage questions |
| Safety Disclaimer Presence | Manual review | Usually present after SFT |
Safety Limitations And Known Failure Cases
- Can hallucinate details or omit important caveats.
- May produce incomplete or awkward sentence endings.
- Can mix nearby concepts (for example fees vs interest vs utilization).
- Not reliable for jurisdiction-specific legal/compliance interpretation.
- Should not be used for credit decisions, underwriting, fraud workflows, or customer-specific advice.
Recommended Safe Usage
- Keep use scope to educational guidance.
- Always verify with official issuer terms and local regulations.
- Add human review before showing output to end users.
Safety Note
Always verify card-specific terms and regulations with the official issuer documentation.
Artifacts
pytorch_model.bin (exported from sft_model.pt)config.jsonchar_tokenizer.json
Example
# Load checkpoint files from this repo and run with the local project code.
