---
language:
- en
license: mit
library_name: transformers
base_model: microsoft/phi-2
base_model_relation: finetune
pipeline_tag: text-generation
tags:
- phi-2
- customer-service
- transcript-analysis
- multi-issue
- legacy
---

# Phi-2 Multi-Issue Transcript Analysis (legacy)

Fine-tune of `microsoft/phi-2` that emits JSON for multi-issue customer-service transcripts (primary/secondary issues, sentiment, agent rating, resolution, CSAT-style score).

**Legacy.** No held-out metric was published. Phi-2 is a 2023 2.7B checkpoint. For process-control labels and forbidden-hit scoring use [Immortalis/cx-comply](https://huggingface.co/datasets/Immortalis/cx-comply). For applied CX decisions with measured accuracy use [chendren/cx-decider-applied-v1](https://huggingface.co/chendren/cx-decider-applied-v1).

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("chendren/phi2-multi-issue-analysis")
tokenizer = AutoTokenizer.from_pretrained("chendren/phi2-multi-issue-analysis")
transcript = "[customer service transcript]"
inputs = tokenizer(transcript, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))
```

## Intended use

Research on structured transcript tagging. Not a production QA scorer. Agent-performance ratings from this model should not be used for HR decisions.

## Limits

- No published test accuracy, ECE, or agreement vs human labels
- English, clear dialogue format
- CSAT values are model guesses, not survey scores

## Citation

```bibtex
@misc{phi2-multi-issue-analysis,
  author = {Hendren, Chad},
  title = {Phi-2 Multi-Issue Transcript Analysis Model},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/chendren/phi2-multi-issue-analysis}
}
```