File size: 3,560 Bytes
3812594 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 |
---
license: llama2
base_model: meta-llama/Llama-2-13b-hf
tags:
- llama2
- computer-engineering
- computer-architecture
- algorithms
- systems
- qora
- lora
- quantized
- merged
language:
- en
pipeline_tag: text-generation
library_name: transformers
datasets:
- cais/mmlu
- sahil2801/CodeAlpaca-20k
- Open-Orca/OpenOrca
model_type: llama
---
# Llama-2-13B-Computer-Engineering
### Overview
**Llama-2-13B-Computer-Engineering** is a fine‑tuned variant of **LLaMA‑2‑13B**, adapted for **computer engineering, computer architecture, systems, and algorithms**.
The model was trained using **QLoRA (4‑bit quantization)**, then merged into a single checkpoint.
This allows **13B‑scale reasoning** to run in ~6.6 GB of storage and ~16GB of GPU memory, making it usable on a single modern GPU.
---
### Training Setup
- **Base model:** [LLaMA‑2‑13B](https://huggingface.co/meta-llama/Llama-2-13b-hf)
- **Fine‑tuning method:** QLoRA (4‑bit NF4) + LoRA adapters (`r=16`, `α=32`)
- **Optimized Layers:** Attention projection modules (`q_proj`, `k_proj`, `v_proj`, `o_proj`)
- **Final merge:** LoRA weights merged into the base model → single merged checkpoint
- **Resulting size:** ~6.6 GB (`safetensors` sharded files) vs. ~24 GB fp16
---
## Dataset
The dataset was curated from multiple sources to emphasize **reasoning, explanations, and code writing in computer engineering contexts**.
Included sources:
- **MMLU (Computer Security subset)** → exam‑style questions on systems and security
- **CodeAlpaca‑20k (filtered)** → algorithm, data structures, complexity, trees, sorting/searching, graphs
- **OpenOrca subset** → reasoning tasks mentioning computer systems and architecture
- **Custom technical examples** (hand‑crafted) on:
- CPU pipelining & instruction‑level parallelism
- Cache coherency and MESI protocol
- Compiler optimizations (instruction scheduling, inlining, loop unrolling)
- RISC vs. CISC architectures
- Memory hierarchies (registers, caches, RAM, storage)
- Branch prediction
- Example algorithms (binary search, stacks, etc.)
---
### Usage
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"Irfanuruchi/Llama-2-13B-Computer-Engineering",
torch_dtype="auto",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Irfanuruchi/Llama-2-13B-Computer-Engineering")
prompt = """### Instruction:
Explain CPU pipelining and its advantages.
### Response:"""
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```
---
## Example Responses
Q: What is cache coherency in multicore systems?
A: Cache coherence ensures that all processor cores observe a consistent view of memory. Protocols such as MESI handle invalidation and updates when one core modifies data, preventing stale values and race conditions.
Q: Implement a stack in Python.
A: Produces a Stack class with methods for push, pop, peek, is_empty, and size.
---
## Limitations
- While optimized for computer engineering, performance outside this scope is similar to the base LLaMA‑2‑13B.
---
## License
- Base model:[Meta's LLaMA 2 license](https://huggingface.co/meta-llama/Llama-2-13b-hf).
- Fine‑tuned weights: Distributed under the same license.
- Datasets: Combination of open academic sets (MMLU, CodeAlpaca, OpenOrca) and custom educational material.
|