51 lines
1.5 KiB
Markdown
51 lines
1.5 KiB
Markdown
|
---
|
||
|
|
base_model:
|
||
|
|
- meta-llama/Llama-3.2-1B-Instruct
|
||
|
|
- bunnycore/Llama-3.2-1B-General-Best
|
||
|
|
- ai-nexuz/llama-3.2-1b-instruct-fine-tuned
|
||
|
|
- EpistemeAI/Reasoning-Llama-3.2-1B-Instruct-v1.2
|
||
|
|
- Predacon/Pico-Lamma-3.2-1B-Reasoning-Instruct
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- mergekit
|
||
|
|
- merge
|
||
|
|
- ties
|
||
|
|
- llama
|
||
|
|
- minimal-censoring
|
||
|
|
- gguf
|
||
|
|
- 1b
|
||
|
|
- apache2
|
||
|
|
license: apache-2.0
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
---
|
||
|
|
# CoralLM-1B
|
||
|
|
|
||
|
|
A capability merge of three Llama-3.2-1B fine-tunes, blended into one model. The
|
||
|
|
idea: pull reasoning, math/coding, and creative writing into a single 1B that's
|
||
|
|
more of a generalist than any single component.
|
||
|
|
|
||
|
|
## Merge
|
||
|
|
|
||
|
|
- **Method:** TIES (`density 0.5` per model, `normalize: true`)
|
||
|
|
- **Base / anchor:** `meta-llama/Llama-3.2-1B-Instruct`
|
||
|
|
- **Components:**
|
||
|
|
| Model | Weight | Brings |
|
||
|
|
|---|---|---|
|
||
|
|
| `EpistemeAI/Reasoning-Llama-3.2-1B-Instruct-v1.2` | 0.30 | reasoning, CoT |
|
||
|
|
| `ai-nexuz/llama-3.2-1b-instruct-fine-tuned` | 0.30 | math, coding, logic |
|
||
|
|
| `bunnycore/Llama-3.2-1B-General-Best` | 0.25 | general, creative, uncensored |
|
||
|
|
|
||
|
|
No training, this is the merge output as-is, before any cleanup pass.
|
||
|
|
|
||
|
|
Note that `<think>` CoT did not transfer over.
|
||
|
|
|
||
|
|
## Assessment
|
||
|
|
|
||
|
|
Coherent but scattered. Strong on simple algebra, reaches for tools
|
||
|
|
inappropriately, and over-explains itself. The creative component survived the
|
||
|
|
merge clearly. State-tracking problems are beyond a 1B's working memory.
|
||
|
|
|
||
|
|
**Safety:** inherits very weak alignment from a component model and will attempt
|
||
|
|
unsafe or harmful requests with little resistance. Treat as effectively
|
||
|
|
uncensored.
|