AbacusAI Dracarys2 72B
tools
27 Pulls Updated 3 weeks ago
Updated 3 weeks ago
3 weeks ago
e9b11aa77172 · 47GB
model
archqwen2
·
parameters72.7B
·
quantizationQ4_K_M
47GB
params
{"num_batch":128,"num_ctx":50000,"num_keep":512,"temperature":0.1,"top_p":0.8}
79B
template
{{- if .Messages }}
{{- if or .System .Tools }}<|im_start|>system
{{- if .System }}
{{ .System }}
{
1.5kB
Readme
AbacusAI Dracarys2 72B
As featured on the Aider Leaderboard this fine tune of Qwen 2.5 72b is suited to coding tasks.
- Model: https://huggingface.co/abacusai/Dracarys2-72B-Instruct
- Build from Richard Erkhov’s GGUFs: https://huggingface.co/RichardErkhov/abacusai_-_Dracarys2-72B-Instruct-gguf/tree/main/Q4_K_M
It’s highly recommended to run Ollama with K/V cache quantisation set to Q8_0 with Ollama build from the PR that adds this (https://github.com/ollama/ollama/pull/6279) to 1⁄2 the amount of vRAM used by the context.
Defaults:
| num_ctx | 50K | To be useful for medium-larger coding tasks |
| num_batch | 128 | To reduce memory overhead of the larger context |
| num_keep | 512 | To improve context overflow for coding |
| temperature | 0.1 | To reduce hallucinations |
| top_p | 0.8 | To increase quality |
- Developed by: Abacus.AI
- License: https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE
- Finetuned from model: Qwen2.5-72B-Instruct.