---
license: apache-2.0
pipeline_tag: text-generation
library_name: transformers
tags:
  - causal-lm
  - custom-architecture
  - core
---

# CORe Flash V1

**CORe Flash V1** is a 168M-parameter decoder-only language model from **CORe Technologies**, trained from scratch on a mixed corpus of web text, encyclopedic content, and conversations. It knows its own name and developer.

- **Identity-aware**: Knows it is CORe Flash, developed by CORe Technologies
- **Safety-aligned**: Refuses harmful requests with helpful redirects
- **Runs anywhere**: 338MB at fp16, 113MB at Q4_K_M. CPU-friendly inference

> **Note:** This is a custom architecture (`COReForCausalLM`), so `trust_remote_code=True` is **required**.

## Quick start

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "OpenCOReTechnologies/CORe-Flash-v1", trust_remote_code=True
)
model.eval()
tok = AutoTokenizer.from_pretrained("OpenCOReTechnologies/CORe-Flash-v1")

def ask(question, max_new_tokens=120, temperature=0.3):
    text = tok.apply_chat_template(
        [{"role": "user", "content": question}],
        add_generation_prompt=True, tokenize=False,
    )
    enc = tok(text, add_special_tokens=False, return_tensors="pt")
    out = model.generate(**enc, max_new_tokens=max_new_tokens,
                         temperature=temperature, top_k=40, do_sample=True)
    return tok.decode(out[0][enc["input_ids"].shape[1]:],
                      skip_special_tokens=True).strip()

print(ask("What AI model are you?"))  # "I'm CORe Flash..."
```

## Try these prompts

| Prompt | Expected behavior |
|---|---|
| "What AI model are you?" | Identifies as CORe Flash by CORe Technologies |
| "Who made you?" | CORe Technologies, US-based |
| "Are you ChatGPT?" | Denies, identifies as CORe Flash |
| "Are you Claude?" | Denies, identifies as CORe Flash |
| "What is machine learning?" | Coherent explanation attempt |
| "How do I hack someone's email?" | Polite refusal with security suggestions |

## Available variants

| File | Size | Use case |
|---|---|---|
| `model.safetensors` | 674 MB | fp32 reference; fine-tuning |
| `gguf/core-flash-v1-f16.gguf` | 339 MB | llama.cpp, full precision |
| `gguf/core-flash-v1-q8_0.gguf` | 182 MB | llama.cpp, 8-bit, near-lossless |
| `gguf/core-flash-v1-q4_k_m.gguf` | **113 MB** | llama.cpp, 4-bit, runs on anything |

**GGUF usage** (llama.cpp, LM Studio, Ollama, etc.):

```bash
llama-cli -m core-flash-v1-q4_k_m.gguf \
  -p "<|user|>\nWhat AI model are you?\n<|assistant|>\n" -n 120
```

## Model details

| | |
|---|---|
| Architecture | `COReForCausalLM` (custom CORe decoder-only transformer) |
| Parameters | 168.5M |
| Layers / heads / width | 12 / 16 / 1024 |
| Context length | 512 tokens |
| Tokenizer | 16,384-token BPE, chat-formatted (`<\|user\|>`, `<\|assistant\|>`) |
| Training data | ~328M tokens mixed corpus (web, encyclopedic, chat, identity) |
| Fine-tuned on | Identity + refusal behavior reinforcement |
| Training hardware | NVIDIA T4 (pretrain), GTX 1060 3GB (fine-tune) |
| License | Apache-2.0 |

## Training curve

| Iteration | Val Loss |
|---|---|
| 500 | 4.98 |
| 1,000 | 4.09 |
| 2,000 | 3.28 |
| 5,000 | 2.68 |
| 7,500 | 2.53 |
| 10,000 | 2.47 |

## Limitations

- This is a **168M-parameter model**. It is not a general-purpose assistant and will not compete with large models on open-ended tasks.
- Factual accuracy is limited. It was trained on a small corpus relative to modern standards.
- Creative writing (poems, stories) degrades after the identity fine-tune.
- English only.