LoopFill experiments: dflash for ouro-2.6b
This repository contains the selected trained drafter used by the LoopFill
baseline experiments. It requires ByteDance/Ouro-2.6B
at revision 1ed04250da1a9936042725d302e81c8fa2ab5abd. Target weights and tokenizer are downloaded
separately from their original repository.
| Property | Value |
|---|---|
| Method | dflash |
| Selected checkpoint step | 9030 |
| Trainable parameters | 277898496 |
| Training epochs / final updates | 6 / 9030 |
| Weight file | model.safetensors |
| SHA256 | 55bf467d26407751d57600a829344a2d02edcaf0a4ba18fe438036e62e7d71ca |
Download and use
from huggingface_hub import snapshot_download
checkpoint = snapshot_download("zsqzz/loopfill-dflash-ouro-2.6b")
Use the custom implementation in zhusq20/loop_suffix
and follow its standalone inference guide.
Pass the downloaded directory as --checkpoint to python run_baseline.py infer
with --method dflash and --model ouro-2.6b. A local matching target model and
JSONL prompts are also required. This artifact is not an independent chat model
or a generic AutoModelForCausalLM checkpoint.
The byte-identical weights, training metadata, selection evidence and source hashes are preserved in this repository. The GitHub experiment archive documents the benchmark inputs, backend, AR denominator, timing boundaries and limitations. LoopAhead is our paper-based implementation, not official author code/weights. DFlash includes our recurrent-target adaptations. Single-seed timing results do not establish cross-seed stability or task-quality equivalence.
Optimizer/RNG states and intermediate training checkpoints are not included. The original target and third-party implementation terms remain applicable; the repository records upstream code license notices where relevant.
- Downloads last month
- 23