# phi2-merged Model Report

## Summary

This folder contains a standalone causal language model with a Phi family architecture:

- `architectures`: `PhiForCausalLM`
- `model_type`: `phi`
- `hidden_size`: `2560`
- `num_hidden_layers`: `32`
- `num_attention_heads`: `32`
- `vocab_size`: `51200`

The weight file is a full model checkpoint, not a lightweight adapter. The safetensors keys are all standard backbone parameters such as `model.layers.*`, `model.embed_tokens.weight`, and `lm_head.*`. There are no LoRA, adapter, prefix-tuning, or chat-template artifacts in the checkpoint layout.

## Base Model or Instruct Model

Best classification from the local evidence: this is a base causal LM, not a clearly packaged instruct/chat model.

Why:

- The config does not declare an instruct or chat variant.
- The tokenizer files do not define a special instruction/chat template.
- The checkpoint layout is a plain full model checkpoint.
- Behavior is mixed: it answers simple prompts and code requests well, but it also repeats or drifts on some Turkish chat-style prompts instead of consistently following a conversational instruction format.

## What It Looks Fine-Tuned For

The model appears strongest in the following areas:

- short factual completions and prompt continuation
- simple arithmetic and reasoning-style prompts
- code generation, especially small Python snippets
- English-language instructions better than Turkish chat formatting

The generated outputs suggest some instruction-following ability, but not the stronger, more stable behavior typical of a dedicated chat-tuned model.

## What It Is Better At

Based on the local probes run in the terminal, the model seems better at:

- direct, narrowly scoped tasks
- code answers with obvious structure
- math-style completions
- continuation after explicit answer cues such as `Answer:` or `Cevap:`

It seems weaker at:

- multi-turn conversational flow
- Turkish dialogue formatting
- avoiding repetition when the prompt is loosely structured

## Evidence From Local Probes

Observed behavior:

- For a Turkish math prompt, it produced a correct `2 + 2 = 4` style answer, but then kept extending into repetitive or mixed reasoning.
- For a Turkish chat prompt, it echoed the prompt content instead of cleanly producing a single assistant reply.
- For an English coding prompt, it produced a clean Python function to reverse a string.

## Recommended Usage

- Use it as a general causal LM or a prompt-completion model.
- Prefer explicit answer cues like `Answer:` or `Cevap:`.
- For chat usage, wrap it with a custom prompt format if you want more stable assistant-style responses.

## Run Command

After activating the upper-level virtual environment, run:

```powershell
& "c:\ai_project\ai_env\Scripts\python.exe" "c:\ai_project\phi2-merged\run_phi2.py" "Kısa bir selam ver:"
```

You can also pass a custom prompt:

```powershell
& "c:\ai_project\ai_env\Scripts\python.exe" "c:\ai_project\phi2-merged\run_phi2.py" "Write a Python function that reverses a string."
```