Skip to content

Instantly share code, notes, and snippets.

View chimezie's full-sized avatar

Chime Ogbuji chimezie

View GitHub Profile
@chimezie
chimezie / axolotl.yaml
Last active May 12, 2025 20:24
Axolotl training on private data
base_model: google/gemma-3-27b-it
#model_type: AutoModelForCausalLM
#tokenizer_type: AutoTokenizer
deepspeed: deepspeed_config.json
load_in_8bit: false
load_in_4bit: true
#strict: false
base_model: google/gemma-2-27b
lora_model_dir: /path/to/previous/axolotl/run/output/
#model_type: AutoModelForCausalLM
#tokenizer_type: AutoTokenizer
save_safetensors: true
load_in_8bit: false
load_in_4bit: true
strict: false
@chimezie
chimezie / gist:6be5752a27745f86b7911df7e0ccb73b
Created December 25, 2024 18:00
A generative, conversational workflow system that uses PDL and mlx_lm for evaluation
"""
Declarative Prompts, Generative program generators, oh my...
Takes a minimal Prompt Declaration Language (PDL) file and generates a finite state generative machine
as Python objects for a subset of the PDL language [1], [2]. These objects (the "programs" in particular)
can be executed, and their Model class can be extended to incorporate the functionality for
evaluating the prompts against the models specified in PDL using any accumulated conversational
context, prompts, and generation parameters (sampling parameters, for example), (optionally) updating
the context as the programs execution continues. A generative, conversational workflow system