Intent Classifier Experiments
This Hugging Face repo stores all experiment-time adapter artifacts for the
intent-classifier project.
In the finetuned models under v1.0 - we were predicting the tool name, while
from v2.0 we are now predicting the tool id.
The v2.1 set is the final artifact set for this repository and is expected to
remain unchanged for the foreseeable future.
Purpose
- Keep every adapter checkpoint produced during fine-tuning.
- Preserve per-experiment traceability across versions.
- Preserve the final v2.1 adapter artifacts alongside the earlier experiments.
- Keep release models separate from this artifact repository.
Layout
v1.0/, v2.0/, and v2.1/ each contain one folder per experiment run.- Folder naming format:
{model}_{technique}_{config}_{dataset_size}_{YYYYMMDD-HHMMSS}
Each experiment folder can include:
- adapter weights
- tokenizer/config files
- metadata and training outputs
v2.0 experiment set
The v2.0 experiment set is complete. It contains 95 runs, all using the 1k
dataset-size setting, produced between 2026-08-25 and 2026-08-27.
- Base models: Qwen 2.5 0.5B, Qwen3 0.6B, Llama 3.2 1B, SmolLM2 360M, and
SmolLM2 1.7B
- Adapter techniques explored: LoRA, DoRA, LoRA+, DoRA+, AdaLoRA, and QLoRA
- Configurations: A through D, where applicable
These are experiment artifacts, not production recommendations. Evaluation,
comparison, and promotion decisions are maintained in the inference/benchmarking
and release repositories below.
v2.1 experiment set
The v2.1 set is final for this repository (as of the foreseeable future).
It contains three runs using the 10k dataset-size setting,
produced between 2026-08-27 and 2026-08-28.
- Base models: Qwen 2.5 0.5B, Qwen3 0.6B, and SmolLM2 360M
- Adapter technique: DoRA+
- Configuration: D
These final artifacts are retained for traceability and reproducibility; model
evaluation and release selection are maintained in the related repositories
below.
Related repositories
Training code and release models are maintained separately.
Versioning
Experiment folders are grouped by release version (for example, v1.0). v2.1 is
the final version in this repository for the foreseeable future. Selected models
are promoted separately to the release repo.
