DCASE paper: training checkpoints

Raw training outputs for the DCASE paper: supervised fine-tuning (SFT) and GRPO runs on top of Qwen/Qwen2-Audio-7B-Instruct.

Layout: <experiment group>/<run>/checkpoint-<step>/ and <experiment group>/<run>/final_model/
Full fine-tunes: 4 bf16 safetensors shards; load with Qwen2AudioForConditionalGeneration.from_pretrained("darkraider42/dcase-paper-checkpoints", subfolder="<group>/<run>/final_model")
LoRA runs: PEFT adapters (r=8, alpha=16, q_proj/v_proj) for the base model
Optimizer, scheduler and RNG state (optimizer.pt, scheduler.pt, rng_state*.pth) are included for resuming training
Runs in debug_* groups are debugging runs

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for darkraider42/dcase-paper-checkpoints

Base model

Adapter

(21)

this model