| --- |
| 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. |
| - **Safety-aligned**: Refuses harmful requests with helpful redirects (not like it would be able to help with harmful requests in the *first* place) |
| - **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` | |
| | 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) | |
| | License | Apache-2.0 | |
|
|
| ## 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) is degraded. |
| - English only. |
