tor-simt-llama-3-8b-gpt4-de-en — git@unknown

README.md

@@ -0,0 +1,45 @@
+---
+tags:
+- simultaneous-translation
+- east
+- ot-annotation
+language: ['de', 'en']
+license: mit
+---
+
+# llama_3_8b_gpt4_de_en
+
+Experiment `llama_3_8b_gpt4_de_en` from the *Teacher-Free Read/Write Annotation for Simultaneous
+Machine Translation* project.
+
+## Recipe
+
+- **Backbone:** `meta-llama/Meta-Llama-3-8B-Instruct`
+- **Corpus:** `SiMT-De-En-660K`
+- **Annotator:** `GPT-4 (shipped with SiMT-De-En-660K, unchanged)`
+- **Criterion:** `gpt4-original` (τ = 0)
+- **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
+
+```bash
+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/llama_3_8b_gpt4_de_en.yaml --ngpus N
+```
+
+Git commit at time of run: `unknown`

chat_template.jinja

@@ -0,0 +1,5 @@
+{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>
+
+'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>
+
+' }}{% endif %}

config.json

@@ -0,0 +1,32 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 128000,
+  "dtype": "bfloat16",
+  "eos_token_id": 128009,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 14336,
+  "max_position_embeddings": 8192,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_parameters": {
+    "rope_theta": 500000.0,
+    "rope_type": "default"
+  },
+  "tie_word_embeddings": false,
+  "transformers_version": "5.14.1",
+  "use_cache": false,
+  "vocab_size": 128258
+}

config.yaml

@@ -0,0 +1,58 @@
+# llama_3_8b_gpt4_de_en — matched-annotator control for the falsifiable claim.
+# Same 10K de-en rows from SiMT-De-En-660K but with the shipped GPT-4
+# source_chunks/target_chunks (EAST's original annotator). Compare head-to-head
+# with llama_3_8b_curated_de_en (10K de-en rows, our OT annotations).
+
+tag: llama_3_8b_gpt4_de_en
+
+backbone:
+  hf_id: meta-llama/Meta-Llama-3-8B-Instruct
+  local_path: ${SIMT_MODEL_BASE}/Meta-Llama-3-8B-Instruct
+  is_instruct: true
+  tokenizer_dir: ${SIMT_MODEL_BASE}/Meta-Llama-3-8B-Instruct-extended-v6
+
+source_pool:
+  corpus: SiMT-De-En-660K
+  directions:
+    de-en: 10000
+  human_target_only: false
+  seed: 42
+
+annotate:
+  annotator: GPT-4 (shipped with SiMT-De-En-660K, unchanged)
+  criterion: gpt4-original
+  tau: 0
+  top_k: 0
+  lookahead_k: 0
+  latency_bins: [low, medium, high]
+
+sft_dataset:
+  merge_small_chunks: false
+  keep_collapsed_rows: false
+
+train:
+  num_epochs: 2
+  per_device_batch_size: 2
+  grad_accum_steps: 8
+  learning_rate: 1.0e-5
+  warmup_steps: 200
+  eval_steps: 500
+  save_steps: 500
+  early_stopping_patience: 3
+  bf16: true
+  completion_only_loss: false
+
+eval:
+  test_sets:
+    wmt15: [de-en]
+    wmt22: [de-en]
+    iwslt17: [de-en]
+  latencies: [low, low-medium, medium, medium-high, high]
+  n_sentences: -1
+  policy: check_argmax
+  mode: streaming
+
+plot:
+  color: "#e6194B"
+  marker: D
+  label: Llama-3-8B w/ GPT-4 chunks

generation_config.json

@@ -0,0 +1,12 @@
+{
+  "bos_token_id": 128000,
+  "do_sample": true,
+  "eos_token_id": [
+    128001,
+    128009
+  ],
+  "max_length": 4096,
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "5.14.1"
+}