llm/SIA
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Replaced deepseek with qwq

Browse Source
This commit is contained in:
Niels Geens
2025-03-14 11:22:30 +01:00
parent 3ea3239a9b
commit 2e66020f8e
13 changed files with 693 additions and 530 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
procedures/self_improvement/reasoning.md
View File

@@ -309,25 +309,21 @@ This preserves the temporal relationships between entries while anchoring them t
## Training Configuration
SIA takes a modular approach to model training by having separate specialized tools for each provider like train_mistral, train_deepseek, etc.
Each tool shares similar core functionality while handling provider-specific requirements.
The default training tool and parameters are called from the `/root/sia/tools/train/train.sh` script.
While the training process is conceptually similar across providers, each has unique requirements for data formatting, API interactions, and job management.
By creating dedicated tools, we can properly encapsulate these differences without complicating the core training logic.
For example, Mistral needs JSONL files with specific message structures, while other providers might require different formats or metadata.
A dedicated `train` tool encapsulates these differences without complicating the surrounding training logic.
Training configuration should be consistent regardless of the provider.
All training tools read from the same config.yaml format, which defines essential parameters like the system prompt, action schema, and training data paths.
Training configuration is consistent regardless of the provider.
The same config.yaml file is supported by all implementations.
This defines essential parameters like the system prompt, action schema, and training data paths.
These parameters represent fundamental aspects of how we want the model to behave, independent of which provider handles the actual training.
The tools then translate these standard parameters into provider-specific settings.
The tool then translates these standard parameters into provider-specific settings for the current active provider.
Training tools enforce important safeguards around version control.
Before starting a training run, each tool verifies that all source files - including the config itself, training data, system prompt, and action schema - are committed to git.
The training tool enforces important safeguards around version control.
Before starting a training run, the tool verifies that all source files are committed to git.
This ensures reproducibility by guaranteeing we can recreate the exact training conditions that produced any given model.
The git commit hash becomes part of the internal tracking of model versions.
The tools follow a common workflow:
The tool follow a common workflow for each provider:
1. Read and validate the standard config.yaml format
2. Check that all source files are committed to git
3. Convert training data into the provider's required format