unswnlporg/tor-simt-llama-3-8b-curated-en-vi

llama38b_curated

Experiment llama_3_8b_curated from the Teacher-Free Read/Write Annotation for Simultaneous Machine Translation project.

Recipe

• —Backbone: meta-llama/Meta-Llama-3-8B-Instruct
• —Corpus: curated
• —Annotator: same_as_backbone
• —Criterion: ot (τ = 0.3)
• —Latencies: ['low', 'medium', 'high']

Files in this repo

• —config.yaml — the exact experiment config that produced this run.
• —manifest.json — git sha, hostname, GPUs, timestamps.
• —logs/ — per-stage stdout+stderr from bin/run.
• —eval/ — every landed eval-JSON cell (hypothesis, reference, AL, BLEU).
• —annotate/ — per-direction matrices.jsonl (divergence matrices) produced by this backbone as annotator.
• —source_pool.json — the corpus rows the matrices index into ({index, source, target, src_lang, tgt_lang, latency, source_chunks, target_chunks, _corpus}). Matrices are unjoinable without this file — records only carry index.
• —SFT checkpoint (*.safetensors + tokenizer).

Reproduce

git clone https://github.com/dipankarsrirag/simt-tor-26.git
cd simt-tor-26
cp .simtrc.example .simtrc  # edit paths for your setup
bin/run configs/01_llama_3_8b_curated.yaml --ngpus N

Git commit at time of run: unknown