Add prepared-data bundle prepared-2026-05-11-smoke32
Browse files- prepared/prepared-2026-05-11-smoke32/config_hash.txt +1 -0
- prepared/prepared-2026-05-11-smoke32/datapool.eval.jsonl +7 -0
- prepared/prepared-2026-05-11-smoke32/datapool.eval.parquet +3 -0
- prepared/prepared-2026-05-11-smoke32/datapool.normalized.jsonl +0 -0
- prepared/prepared-2026-05-11-smoke32/datapool.normalized.parquet +3 -0
- prepared/prepared-2026-05-11-smoke32/datapool.train.jsonl +0 -0
- prepared/prepared-2026-05-11-smoke32/datapool.train.parquet +3 -0
- prepared/prepared-2026-05-11-smoke32/datapool.train.sampled.jsonl +0 -0
- prepared/prepared-2026-05-11-smoke32/datapool.train.sampled.parquet +3 -0
- prepared/prepared-2026-05-11-smoke32/effective_config.yaml +381 -0
- prepared/prepared-2026-05-11-smoke32/ood_test.jsonl +0 -0
- prepared/prepared-2026-05-11-smoke32/prepare_data_summary.json +26 -0
- prepared/prepared-2026-05-11-smoke32/prepared_manifest.json +96 -0
prepared/prepared-2026-05-11-smoke32/config_hash.txt
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
ee9eaf491a503a5f42dc8688119e1f4753091dfa2e69be59110582d3623791d4
|
prepared/prepared-2026-05-11-smoke32/datapool.eval.jsonl
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"dataset":"alwaysgood/reuter-semantic-deduped","id":"alwaysgood/reuter-semantic-deduped:00000019","metadata":{"chunk_idx":null,"document_type":"article","original_id":"alwaysgood/reuter-semantic-deduped:00000019","parent_id":null,"text_role":"body","title":"BP's CEO shake-up could pave way to mega merger"},"source":"Bp Plc Chevron Corp Exxon Mobil Corp LONDON, Dec 18 (Reuters) - The surprise appointment of Meg O'Neill as BP's first outsider CEO offers the bruised $90 billion British oil company three clear strategic choices for moving forward: build, buy or be bought. BP announced on Wednesday that CEO Murray Auchincloss would step down the following day and that O'Neill, current CEO of Australian-listed Woodside Energy , would succeed him in April. BP's head of trading Carol Howle will act as interim CEO until then. The leadership overhaul is the clearest signal yet that BP Chair Albert Manifold, who himself took office on October 1, intends to fundamentally overhaul the long-struggling oil company. The shake-up follows three tumultuous years for the 116-year-old energy firm that saw leadership changes, the overhaul of a failed renewables strategy, activist investor pressure and swirling rumours that the company would soon be acquired. \"Progress has been made in recent years, but increased rigor and diligence are required to make the necessary transformative changes to maximise value for our shareholders,” Manifold said in a statement on Wednesday. A DREAM JOB? O'Neill, a U.S. citizen, will not only become BP's first outsider CEO but also the first woman to lead a Western oil major. She oversaw Woodside's expansion in recent years into a larger player in the liquefied natural gas space. Taking the helm of BP is a job on an entirely different scale, however. Auchincloss became CEO in January 2024 following the shock resignation of Bernard Looney in September 2023 for lying to the board about personal relationships with employees. Auchincloss radically reversed his predecessor's ambitious energy transition strategy, scrapping plans to reduce BP's oil and gas production and expand its renewables business. Following heavy pressure from activist investor Elliott Management, Auchincloss in February vowed to further focus BP's operations on its legacy business as returns from renewables languished. BP has revived its oil and gas division with several new upstream projects in the Gulf of Mexico, Iraq, and Libya as well as a major oil discovery in Brazil . And the company, which pumped 2.36 million barrels of oil equivalent per day (boepd) in 2024, now aims to bump that up slightly to between 2.3 million and 2.5 million boepd by 2030. BP also launched a $4 billion to $5 billion cost-cutting drive by 2027 and this year cut 5% of its global workforce of nearly 90,000. These changes appear to have gone down well with investors. BP's shares are up 10% so far this year, outperforming Shell , Chevron and TotalEnergies . O'Neill thus inherits a company moving in the right direction. But she will now be tasked with further improving BP's operational and financial performance under the supervision of Manifold, who seems to be far more involved in the daily running of the company than his recent predecessors. The safest bet for O’Neill would probably be to continue growing BP's upstream division by diverting more cash to the business to develop production or acquire assets. The company currently plans to spend around $10 billion in oil and gas out of a total capital expenditure budget of $13 billion to $15 billion per year between 2025 and 2027. She will also likely stick to or go beyond her predecessor's plans to make $20 billion in divestments between 2025 and 2027 in order to slash BP’s heavy debt burden. Forecasts of weakening oil prices next year amid surging global supplies will complicate O'Neill's task, however. A line graph showing that BP's price return has underperformed Exxon, Chevron, Shell and TotalEnergies since roughly 2021 BETTER IS RISKIER Ironically, the better BP performs in the coming year, the more uncertain its future becomes. Over the past few years, there has been much speculation that BP will be acquired by a bigger rival. Shell in June ruled out purchasing BP following media reports that it was planning to do so. Other companies including Exxon Mobil , Chevron and the United Arab Emirates' national oil company Adnoc have reportedly considered bidding for BP. To be sure, acquiring the British energy major would be a hugely costly and complex transaction that would require lengthy regulatory approvals and massive disposals. Integrating the firm would also likely take years, with many bumps along the way. But with the industry’s growing confidence that oil and gas demand will remain robust for decades despite the energy transition, consolidation has become more attractive, as it offers companies an opportunity to grow scale and reduce operational costs. The recent growth and success of BP's upstream portfolio therefore makes it an attractive prospect for a rival – such as Shell – seeking to grow its near and long-term oil and gas production. The appointment of O’Neill is a signal that BP’s board wants to keep fighting for the company’s future after three bruising years. Their success may nevertheless place it squarely in the sights of acquisitive rivals. Want to receive my column in your inbox every Monday and Thursday, along with additional energy insights and links to trending stories? Sign up for my Power Up newsletter here . Enjoying this column? Check out Reuters Open Interest (ROI), your essential new source for global financial commentary. ROI delivers thought-provoking, data-driven analysis. Markets are moving faster than ever. ROI can help you keep up. Follow ROI on LinkedIn and X. Ron Bousso; Editing by Jamie Freed"}
|
| 2 |
+
{"dataset":"alwaysgood/c4-semantic-deduped","id":"alwaysgood/c4-semantic-deduped:00000037","metadata":{"chunk_idx":null,"document_type":"other","original_id":"alwaysgood/c4-semantic-deduped:00000037","parent_id":null,"text_role":"body","title":null},"source":"A celebration of the strength and resilience and femininity of women itself. An homage to all woman, throughout a vision of power and charismatic women that turned the course of history. Marie Curie, Frida Kahlo, Angela Davis, Winnie Mandela, Audrey Hepburn, Twiggy, Grace Jones and Janis Joplin are just some of great woman that changed the world in their own way. Let’s celebrate their great achievements on behalf of all women! WOMAN - Who runs the world? Political activists, scientists, artists and leaders in their own way. There is something these inspirational women all share in common: they are all warriors and continue to inspire us in our own modern lives. Dedicated models designed to be a tribute to 4 great female figures from history, from 4 different areas, representing all womanhood. Plain colours and a mix between prints and embroideries T’s, are complemented with floral patterned top & bottom outfits. UNISEX - Gender equality, time to (R)Evolution! Gender equality is achieved when women and men enjoy the same rights and opportunities across all sectors of society, including economic participation and decision-making, and when the different behaviours, aspirations and needs of women and men are equally valued. Why can’t they dress the same? Floral prints and revolutionary messages for both man and woman. Blue, cherry and beige are the statement colours. MAN - Real Men are Feminists! Being feminist is a demonstration of strength and recognition of equal rights, negating domination of all kinds. Strong messages, a large pallet of colours from violet to green and digital printed photos are the highlights for men collection. Watching over the environment and for the comfort of our collections, most of our garments are in natural fabrics. Organic fabrics have high absorption quality, the reason why wearing organic garments can be very comfortable even on hotter days. Plus, are naturally hypoallergenic and have unique anti-bacterial properties."}
|
| 3 |
+
{"dataset":"alwaysgood/c4-semantic-deduped","id":"alwaysgood/c4-semantic-deduped:00000060","metadata":{"chunk_idx":null,"document_type":"other","original_id":"alwaysgood/c4-semantic-deduped:00000060","parent_id":null,"text_role":"body","title":null},"source":"One day minding my own business i thought of my ex boyfriend the true love of my life and there on the ground was a feather. Very interesting to read about feathers. I just got out of jail last week and have had a recent huge breakup in a relationship but know its nest for my mental health. She is the older sister of finn and. White feathers appearing out of nowhere. In addition here is a quick look at some of the additional meaning certain colored feathers may have. Ravenna is the cruel and tyrannical ruler of the kingdom of tabor as well as a dangerously powerful sorceress. Ravenna also known asthe evil queen or simplythe queen is the primary antagonist of snow white and the huntsman and the huntsman. I asked back in december 09 nothing. Pink feathers are a reminder of the unconditional love of your angels and a reminder of the infinite inspiration available when youre tuned into love. I cannot believe all the feathers that have appeared to me in the last 2 months."}
|
| 4 |
+
{"dataset":"alwaysgood/c4-semantic-deduped","id":"alwaysgood/c4-semantic-deduped:00000062","metadata":{"chunk_idx":null,"document_type":"other","original_id":"alwaysgood/c4-semantic-deduped:00000062","parent_id":null,"text_role":"body","title":null},"source":"THE STORY: Bob was born in Fort Cobb, Oklahoma May 6, 1930. Bob loved to tell stories about his life, his family and friends and was able to share these stories with them during his last few days. Bob graduated from Ripon High School where he was a basketball star. He was a fan of local sports teams, especially Central Catholic and Modesto Christian Schools. Honorably discharged from the U.S. Navy in 1954; he served 4 years during the Korean War aboard the Battleship USS Missouri. Bob and his beautiful wife Connie owned Modesto Coin Center, where they worked together for 33 years. Bob was life member #1 of the Stanislaus County Coin Club, served as President, and attended many coin show throughout California. Bob passed away on Jan. 29, 2008 at age 77."}
|
| 5 |
+
{"dataset":"alwaysgood/sec-10k-semantic-deduped","id":"alwaysgood/sec-10k-semantic-deduped:00000014","metadata":{"chunk_idx":null,"document_type":"filing","original_id":"alwaysgood/sec-10k-semantic-deduped:00000014","parent_id":null,"text_role":"body","title":null},"source":"The Company employs approximately 35,000 persons."}
|
| 6 |
+
{"dataset":"alwaysgood/sec-10k-semantic-deduped","id":"alwaysgood/sec-10k-semantic-deduped:00000035","metadata":{"chunk_idx":null,"document_type":"filing","original_id":"alwaysgood/sec-10k-semantic-deduped:00000035","parent_id":null,"text_role":"body","title":null},"source":"Additional information regarding the Partnership is set forth in the Partnership's Annual Report on Form 10-K (Commission File No. 1-9433), which is separately filed with the Securities and Exchange Commission (the \"Commission\")."}
|
| 7 |
+
{"dataset":"alwaysgood/sec-10k-semantic-deduped","id":"alwaysgood/sec-10k-semantic-deduped:00000042","metadata":{"chunk_idx":null,"document_type":"filing","original_id":"alwaysgood/sec-10k-semantic-deduped:00000042","parent_id":null,"text_role":"body","title":null},"source":"The Company intends to continue to expand its organic chemicals business by developing and/or distributing new products that draw upon its skills in organic chemical synthesis and/or its sales organization."}
|
prepared/prepared-2026-05-11-smoke32/datapool.eval.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2d77d52f500e71a21a7af9609c0b39017a7af3362c3ce3bd4790176df43eb6da
|
| 3 |
+
size 12660
|
prepared/prepared-2026-05-11-smoke32/datapool.normalized.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
prepared/prepared-2026-05-11-smoke32/datapool.normalized.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8e03989eb6ab54e41818f2dab11a37604ef58e93549742254c4ac66b8eda36be
|
| 3 |
+
size 239240
|
prepared/prepared-2026-05-11-smoke32/datapool.train.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
prepared/prepared-2026-05-11-smoke32/datapool.train.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:beb0f3ebc295f967dc61dc34628bf20b5fff93de491b6fabe610e7fb47e46210
|
| 3 |
+
size 235090
|
prepared/prepared-2026-05-11-smoke32/datapool.train.sampled.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
prepared/prepared-2026-05-11-smoke32/datapool.train.sampled.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:beb0f3ebc295f967dc61dc34628bf20b5fff93de491b6fabe610e7fb47e46210
|
| 3 |
+
size 235090
|
prepared/prepared-2026-05-11-smoke32/effective_config.yaml
ADDED
|
@@ -0,0 +1,381 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
model:
|
| 2 |
+
name: alwaysgood/qwen35-it
|
| 3 |
+
load_in_4bit: false
|
| 4 |
+
max_length: 8192
|
| 5 |
+
max_seq_length: 8192
|
| 6 |
+
dtype: bf16
|
| 7 |
+
attention_impl: xformers
|
| 8 |
+
padding_side: left
|
| 9 |
+
eos_token: null
|
| 10 |
+
chat_template: null
|
| 11 |
+
trust_remote_code: false
|
| 12 |
+
data:
|
| 13 |
+
runtime:
|
| 14 |
+
mode: hf
|
| 15 |
+
local_jsonl_path: null
|
| 16 |
+
prepare_data:
|
| 17 |
+
intermediate_format: parquet
|
| 18 |
+
parquet_row_group_size: 4096
|
| 19 |
+
progress_enabled: true
|
| 20 |
+
progress_every_rows: 100000
|
| 21 |
+
progress_every_seconds: 10.0
|
| 22 |
+
hf:
|
| 23 |
+
streaming: false
|
| 24 |
+
max_rows_per_dataset: 64
|
| 25 |
+
dataset_download_workers: 2
|
| 26 |
+
fallback_to_snapshot_jsonl: true
|
| 27 |
+
datasets:
|
| 28 |
+
- name: alwaysgood/reuter-semantic-deduped
|
| 29 |
+
split: train
|
| 30 |
+
data_files:
|
| 31 |
+
train: deduped_shards/deduped_*.parquet
|
| 32 |
+
- name: alwaysgood/c4-semantic-deduped
|
| 33 |
+
split: train
|
| 34 |
+
data_files:
|
| 35 |
+
train: deduped_shards/deduped_*.parquet
|
| 36 |
+
- name: alwaysgood/bloomberg-semantic-deduped
|
| 37 |
+
split: train
|
| 38 |
+
data_files:
|
| 39 |
+
train: deduped_shards/deduped_*.parquet
|
| 40 |
+
- name: alwaysgood/earnings-semantic-deduped
|
| 41 |
+
split: train
|
| 42 |
+
data_files:
|
| 43 |
+
train: deduped_shards/deduped_*.parquet
|
| 44 |
+
- name: alwaysgood/sec-10k-semantic-deduped
|
| 45 |
+
split: train
|
| 46 |
+
data_files:
|
| 47 |
+
train: deduped_shards/deduped_*.parquet
|
| 48 |
+
text_columns:
|
| 49 |
+
- source_text
|
| 50 |
+
- text
|
| 51 |
+
subset_size: 32
|
| 52 |
+
split:
|
| 53 |
+
eval_ratio: 0.02
|
| 54 |
+
seed: 42
|
| 55 |
+
sampling:
|
| 56 |
+
strategy: first_n
|
| 57 |
+
seed: 42
|
| 58 |
+
length:
|
| 59 |
+
enabled: true
|
| 60 |
+
mode: tokenizer
|
| 61 |
+
max_total_tokens: 8192
|
| 62 |
+
max_source_tokens: 3900
|
| 63 |
+
max_output_tokens: 4096
|
| 64 |
+
prompt_template_tokens: 256
|
| 65 |
+
min_available_output_tokens: 768
|
| 66 |
+
safety_margin_tokens: 64
|
| 67 |
+
overflow: split
|
| 68 |
+
output_budget_strategy: dynamic
|
| 69 |
+
tokenizer_name: null
|
| 70 |
+
tokenizer_fallback: whitespace
|
| 71 |
+
tokenizer_local_files_only: false
|
| 72 |
+
tokenizer_batch_size: 16384
|
| 73 |
+
split:
|
| 74 |
+
unit: sentence
|
| 75 |
+
max_source_tokens_per_chunk: 3900
|
| 76 |
+
max_chunks_per_row: 4
|
| 77 |
+
min_chunk_tokens: 32
|
| 78 |
+
fallback_for_long_sentence: skip
|
| 79 |
+
on_max_chunks_exceeded: skip
|
| 80 |
+
ood_test:
|
| 81 |
+
enabled: true
|
| 82 |
+
path: data/test.csv
|
| 83 |
+
source_column: Source_En
|
| 84 |
+
target_column: Target_Ko
|
| 85 |
+
num_workers: 4
|
| 86 |
+
inference:
|
| 87 |
+
runtime:
|
| 88 |
+
mode: subprocess
|
| 89 |
+
unsloth:
|
| 90 |
+
enabled: true
|
| 91 |
+
fallback_to_transformers: false
|
| 92 |
+
multi_gpu:
|
| 93 |
+
enabled: true
|
| 94 |
+
gpu_ids:
|
| 95 |
+
- 0
|
| 96 |
+
- 1
|
| 97 |
+
- 2
|
| 98 |
+
- 3
|
| 99 |
+
- 4
|
| 100 |
+
- 5
|
| 101 |
+
- 6
|
| 102 |
+
- 7
|
| 103 |
+
shard_strategy: order_split
|
| 104 |
+
subprocess:
|
| 105 |
+
command:
|
| 106 |
+
- python3
|
| 107 |
+
- -m
|
| 108 |
+
- scp_stage4.pipeline.workers.vllm_inference_worker
|
| 109 |
+
q1:
|
| 110 |
+
do_sample: true
|
| 111 |
+
temperature: 1.1
|
| 112 |
+
top_p: 0.95
|
| 113 |
+
max_new_tokens: 4096
|
| 114 |
+
on_truncation: discard
|
| 115 |
+
q2:
|
| 116 |
+
do_sample: false
|
| 117 |
+
temperature: 0.0
|
| 118 |
+
top_p: null
|
| 119 |
+
max_new_tokens: 4096
|
| 120 |
+
on_truncation: discard
|
| 121 |
+
throughput:
|
| 122 |
+
batching:
|
| 123 |
+
strategy: token_budget
|
| 124 |
+
max_batch_tokens: 65536
|
| 125 |
+
pad_to_multiple_of: 8
|
| 126 |
+
preserve_order: false
|
| 127 |
+
restore_order_in_artifacts: true
|
| 128 |
+
pipeline:
|
| 129 |
+
subset:
|
| 130 |
+
strategy: fraction
|
| 131 |
+
fraction: 0.02
|
| 132 |
+
min_size: 32
|
| 133 |
+
max_size: null
|
| 134 |
+
shuffle: false
|
| 135 |
+
seed: 42
|
| 136 |
+
drop_last: false
|
| 137 |
+
stage:
|
| 138 |
+
max_subsets: null
|
| 139 |
+
use_sampled_data: false
|
| 140 |
+
subset_archive:
|
| 141 |
+
enabled: false
|
| 142 |
+
format: tar.gz
|
| 143 |
+
output_dir: archives/subsets
|
| 144 |
+
delete_original_after_archive: false
|
| 145 |
+
execution:
|
| 146 |
+
allow_microbatch_overlap: true
|
| 147 |
+
allow_next_subset_q1_prefetch: true
|
| 148 |
+
api_concurrency_overlaps_gpu_work: true
|
| 149 |
+
eval_after_subset:
|
| 150 |
+
enabled: true
|
| 151 |
+
dataset: ood_test
|
| 152 |
+
every_n_subsets: 1
|
| 153 |
+
run_on_final_subset: true
|
| 154 |
+
runtime: qe_subprocess
|
| 155 |
+
source_column: Source_En
|
| 156 |
+
reference_column: Target_Ko
|
| 157 |
+
metrics:
|
| 158 |
+
- xcomet
|
| 159 |
+
- BLEU
|
| 160 |
+
- chrF
|
| 161 |
+
training:
|
| 162 |
+
backend: unsloth
|
| 163 |
+
runtime:
|
| 164 |
+
mode: subprocess
|
| 165 |
+
subprocess:
|
| 166 |
+
collapse_command:
|
| 167 |
+
- torchrun
|
| 168 |
+
- --nproc_per_node=8
|
| 169 |
+
- -m
|
| 170 |
+
- scp_stage4.pipeline.workers.training_worker
|
| 171 |
+
unload_command:
|
| 172 |
+
- torchrun
|
| 173 |
+
- --nproc_per_node=1
|
| 174 |
+
- -m
|
| 175 |
+
- scp_stage4.pipeline.workers.training_worker
|
| 176 |
+
update_command:
|
| 177 |
+
- torchrun
|
| 178 |
+
- --nproc_per_node=8
|
| 179 |
+
- -m
|
| 180 |
+
- scp_stage4.pipeline.workers.training_worker
|
| 181 |
+
collapse_lora:
|
| 182 |
+
rank: 4
|
| 183 |
+
learning_rate: 0.0025
|
| 184 |
+
num_train_epochs: 1
|
| 185 |
+
dropout: 0.0
|
| 186 |
+
bias: none
|
| 187 |
+
response_template: '### Response:
|
| 188 |
+
|
| 189 |
+
'
|
| 190 |
+
batching:
|
| 191 |
+
per_device_train_batch_size: 64
|
| 192 |
+
gradient_accumulation_steps: 1
|
| 193 |
+
packing: false
|
| 194 |
+
base_update:
|
| 195 |
+
mode: full_weight
|
| 196 |
+
persistence: cumulative
|
| 197 |
+
save_after_each_subset: true
|
| 198 |
+
num_train_epochs: 1
|
| 199 |
+
lora:
|
| 200 |
+
rank: 32
|
| 201 |
+
alpha: 64
|
| 202 |
+
dropout: 0.0
|
| 203 |
+
bias: none
|
| 204 |
+
use_rslora: false
|
| 205 |
+
loftq_config: null
|
| 206 |
+
target_modules:
|
| 207 |
+
- q_proj
|
| 208 |
+
- k_proj
|
| 209 |
+
- v_proj
|
| 210 |
+
- o_proj
|
| 211 |
+
- gate_proj
|
| 212 |
+
- up_proj
|
| 213 |
+
- down_proj
|
| 214 |
+
- in_proj_a
|
| 215 |
+
- in_proj_b
|
| 216 |
+
- in_proj_z
|
| 217 |
+
- in_proj_qkv
|
| 218 |
+
- out_proj
|
| 219 |
+
optimizer:
|
| 220 |
+
learning_rate: 1.0e-05
|
| 221 |
+
weight_decay: 0.0
|
| 222 |
+
warmup_ratio: 0.1
|
| 223 |
+
lr_scheduler_type: cosine
|
| 224 |
+
optim: adamw_torch
|
| 225 |
+
max_grad_norm: 1.0
|
| 226 |
+
batching:
|
| 227 |
+
per_device_train_batch_size: 8
|
| 228 |
+
gradient_accumulation_steps: 4
|
| 229 |
+
per_device_eval_batch_size: 8
|
| 230 |
+
packing: false
|
| 231 |
+
response_template: '### Response:
|
| 232 |
+
|
| 233 |
+
'
|
| 234 |
+
data:
|
| 235 |
+
use_no_change: true
|
| 236 |
+
checkpoint:
|
| 237 |
+
save_after_each_subset: true
|
| 238 |
+
save_latest_pointer: true
|
| 239 |
+
keep_subset_checkpoints: true
|
| 240 |
+
keep_last_n: 1
|
| 241 |
+
keep_best_n: 1
|
| 242 |
+
metric_for_best: ood/xcomet_mean
|
| 243 |
+
greater_is_better: true
|
| 244 |
+
keep_final: true
|
| 245 |
+
save_optimizer_state: true
|
| 246 |
+
save_collapse_lora: false
|
| 247 |
+
upload_to_wandb: false
|
| 248 |
+
qe:
|
| 249 |
+
runtime:
|
| 250 |
+
mode: subprocess
|
| 251 |
+
subprocess:
|
| 252 |
+
command:
|
| 253 |
+
- python3
|
| 254 |
+
- -m
|
| 255 |
+
- scp_stage4.pipeline.workers.qe_worker
|
| 256 |
+
multi_gpu:
|
| 257 |
+
enabled: true
|
| 258 |
+
gpu_ids:
|
| 259 |
+
- 0
|
| 260 |
+
- 1
|
| 261 |
+
- 2
|
| 262 |
+
- 3
|
| 263 |
+
- 4
|
| 264 |
+
- 5
|
| 265 |
+
- 6
|
| 266 |
+
- 7
|
| 267 |
+
epsilon: 1.0e-06
|
| 268 |
+
primary:
|
| 269 |
+
backend: comet_kiwi
|
| 270 |
+
model_name: Unbabel/wmt23-cometkiwi-da-xl
|
| 271 |
+
tokenizer_name: null
|
| 272 |
+
batch_size: 64
|
| 273 |
+
max_input_length: 1536
|
| 274 |
+
score_direction: higher_is_better
|
| 275 |
+
transform:
|
| 276 |
+
type: none
|
| 277 |
+
max_score: 1.0
|
| 278 |
+
clamp_for_quality: false
|
| 279 |
+
preserve_raw_score: true
|
| 280 |
+
scoring:
|
| 281 |
+
weighted_score:
|
| 282 |
+
enabled: true
|
| 283 |
+
alpha: 0.15
|
| 284 |
+
beta: 0.85
|
| 285 |
+
selection:
|
| 286 |
+
default_rule:
|
| 287 |
+
require_score_s_gte: 0.0
|
| 288 |
+
top_fraction: 0.01
|
| 289 |
+
isolation:
|
| 290 |
+
enabled: true
|
| 291 |
+
env:
|
| 292 |
+
comet_python_env: COMET_PYTHON
|
| 293 |
+
metricx_python_env: METRICX_PYTHON
|
| 294 |
+
external_api:
|
| 295 |
+
runtime:
|
| 296 |
+
mode: subprocess
|
| 297 |
+
subprocess:
|
| 298 |
+
command:
|
| 299 |
+
- python3
|
| 300 |
+
- -m
|
| 301 |
+
- scp_stage4.pipeline.workers.external_api_worker
|
| 302 |
+
routing:
|
| 303 |
+
mode: single
|
| 304 |
+
fallback_enabled: false
|
| 305 |
+
ensemble_enabled: false
|
| 306 |
+
primary:
|
| 307 |
+
provider: openai
|
| 308 |
+
model: gpt-5.4
|
| 309 |
+
api_key_env: OPENAI_API_KEY
|
| 310 |
+
providers:
|
| 311 |
+
openai:
|
| 312 |
+
enabled: true
|
| 313 |
+
api_key_env: OPENAI_API_KEY
|
| 314 |
+
gemini:
|
| 315 |
+
enabled: false
|
| 316 |
+
api_key_env: GEMINI_API_KEY
|
| 317 |
+
claude:
|
| 318 |
+
enabled: false
|
| 319 |
+
api_key_env: ANTHROPIC_API_KEY
|
| 320 |
+
qwen:
|
| 321 |
+
enabled: false
|
| 322 |
+
api_key_env: QWEN_API_KEY
|
| 323 |
+
deepseek:
|
| 324 |
+
enabled: false
|
| 325 |
+
api_key_env: DEEPSEEK_API_KEY
|
| 326 |
+
failure:
|
| 327 |
+
retry: 2
|
| 328 |
+
on_error: skip
|
| 329 |
+
max_failure_rate: 0.2
|
| 330 |
+
on_failure_rate_exceeded: fail
|
| 331 |
+
output_status:
|
| 332 |
+
allowed:
|
| 333 |
+
- ok
|
| 334 |
+
- skipped
|
| 335 |
+
- failed
|
| 336 |
+
- filtered
|
| 337 |
+
- needs_review
|
| 338 |
+
logging:
|
| 339 |
+
report_to:
|
| 340 |
+
- wandb
|
| 341 |
+
run_name: null
|
| 342 |
+
required_event_fields:
|
| 343 |
+
- run_id
|
| 344 |
+
- subset_idx
|
| 345 |
+
- phase
|
| 346 |
+
- config_hash
|
| 347 |
+
local:
|
| 348 |
+
enabled: true
|
| 349 |
+
root_dir: artifacts/runs
|
| 350 |
+
write_effective_config: true
|
| 351 |
+
write_config_hash: true
|
| 352 |
+
events_jsonl: events.jsonl
|
| 353 |
+
metrics_jsonl: metrics.jsonl
|
| 354 |
+
failures_jsonl: failures.jsonl
|
| 355 |
+
wandb:
|
| 356 |
+
enabled: true
|
| 357 |
+
project: scp_main
|
| 358 |
+
entity: null
|
| 359 |
+
tags:
|
| 360 |
+
- scp
|
| 361 |
+
- stage4
|
| 362 |
+
- sft
|
| 363 |
+
- en-ko
|
| 364 |
+
notes: ''
|
| 365 |
+
log_artifacts: true
|
| 366 |
+
log_model_checkpoints: false
|
| 367 |
+
weave:
|
| 368 |
+
enabled: true
|
| 369 |
+
project: scp_main
|
| 370 |
+
print_call_link: false
|
| 371 |
+
implicitly_patch_integrations: true
|
| 372 |
+
trace_external_api: true
|
| 373 |
+
trace_qe_subprocess: false
|
| 374 |
+
redact_inputs: false
|
| 375 |
+
prompts:
|
| 376 |
+
version: teacher_correction_v1
|
| 377 |
+
preserve_source_in_parentheses: false
|
| 378 |
+
run:
|
| 379 |
+
run_id: real_runtime
|
| 380 |
+
seed: 42
|
| 381 |
+
version: 1
|
prepared/prepared-2026-05-11-smoke32/ood_test.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
prepared/prepared-2026-05-11-smoke32/prepare_data_summary.json
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"artifact_dir": "artifacts/data",
|
| 3 |
+
"cache_hit": false,
|
| 4 |
+
"data_config_hash": "f41f6c5374d68b1a5f52c32eef23bb02bfc6c1eb89ee4c329a6f44abbd226dcb",
|
| 5 |
+
"eval_parquet_rows": 7,
|
| 6 |
+
"eval_rows": 7,
|
| 7 |
+
"intermediate_format": "parquet",
|
| 8 |
+
"length_policy": {
|
| 9 |
+
"input_rows": 320,
|
| 10 |
+
"output_rows": 319,
|
| 11 |
+
"skipped_long_sentence": 2,
|
| 12 |
+
"skipped_max_chunks_exceeded": 0,
|
| 13 |
+
"skipped_overflow_policy": 0,
|
| 14 |
+
"skipped_split_empty": 0,
|
| 15 |
+
"skipped_total": 2,
|
| 16 |
+
"skipped_truncate_budget": 0,
|
| 17 |
+
"split_input_rows": 3,
|
| 18 |
+
"split_output_rows": 2
|
| 19 |
+
},
|
| 20 |
+
"normalized_rows": 319,
|
| 21 |
+
"progress_enabled": true,
|
| 22 |
+
"sampled_parquet_rows": 312,
|
| 23 |
+
"sampled_rows": 32,
|
| 24 |
+
"train_parquet_rows": 312,
|
| 25 |
+
"train_rows": 312
|
| 26 |
+
}
|
prepared/prepared-2026-05-11-smoke32/prepared_manifest.json
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bundle_tag": "prepared-2026-05-11-smoke32",
|
| 3 |
+
"config_hash": "ee9eaf491a503a5f42dc8688119e1f4753091dfa2e69be59110582d3623791d4",
|
| 4 |
+
"config_overrides": [
|
| 5 |
+
"data.runtime.hf.max_rows_per_dataset=64",
|
| 6 |
+
"data.subset_size=32",
|
| 7 |
+
"pipeline.subset.shuffle=false",
|
| 8 |
+
"data.length.tokenizer_fallback=whitespace"
|
| 9 |
+
],
|
| 10 |
+
"config_path": "configs/scp_stage4_real.yaml",
|
| 11 |
+
"created_at_utc": "2026-05-12T04:09:31.352679+00:00",
|
| 12 |
+
"files": [
|
| 13 |
+
{
|
| 14 |
+
"format": "jsonl",
|
| 15 |
+
"path": "datapool.eval.jsonl",
|
| 16 |
+
"row_count": 7,
|
| 17 |
+
"row_id_sha256": "dc33cb57663d9dea11409bbaa70b5176cb9aad22a466be8820fd661ec8dd5c3c",
|
| 18 |
+
"sha256": "299ec305f3f5c0047098a266a166c4975c79069b2dae89a0136478fa49b163fb",
|
| 19 |
+
"size_bytes": 11706
|
| 20 |
+
},
|
| 21 |
+
{
|
| 22 |
+
"format": "parquet",
|
| 23 |
+
"path": "datapool.eval.parquet",
|
| 24 |
+
"row_count": 7,
|
| 25 |
+
"row_id_sha256": "dc33cb57663d9dea11409bbaa70b5176cb9aad22a466be8820fd661ec8dd5c3c",
|
| 26 |
+
"sha256": "2d77d52f500e71a21a7af9609c0b39017a7af3362c3ce3bd4790176df43eb6da",
|
| 27 |
+
"size_bytes": 12660
|
| 28 |
+
},
|
| 29 |
+
{
|
| 30 |
+
"format": "jsonl",
|
| 31 |
+
"path": "datapool.normalized.jsonl",
|
| 32 |
+
"row_count": 319,
|
| 33 |
+
"row_id_sha256": "9efe767cea6072101d49125769368132fd67ad0183eea33380845e5ea3aabfbb",
|
| 34 |
+
"sha256": "73232b4f4b0c5a59e8dfdba43897b5929dd209fa570ab4a4ce2271a126746320",
|
| 35 |
+
"size_bytes": 641669
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"format": "parquet",
|
| 39 |
+
"path": "datapool.normalized.parquet",
|
| 40 |
+
"row_count": 319,
|
| 41 |
+
"row_id_sha256": "9efe767cea6072101d49125769368132fd67ad0183eea33380845e5ea3aabfbb",
|
| 42 |
+
"sha256": "8e03989eb6ab54e41818f2dab11a37604ef58e93549742254c4ac66b8eda36be",
|
| 43 |
+
"size_bytes": 239240
|
| 44 |
+
},
|
| 45 |
+
{
|
| 46 |
+
"format": "jsonl",
|
| 47 |
+
"path": "datapool.train.jsonl",
|
| 48 |
+
"row_count": 312,
|
| 49 |
+
"row_id_sha256": "2571fe9e2e8b5e73e4b3705147ace20409c40679ff959a1f94d6103f6b6f1df6",
|
| 50 |
+
"sha256": "d2244e1dedda59d4f5d7399a6be0beac933a3d28c30e799de0dedcef3b6ff2f0",
|
| 51 |
+
"size_bytes": 629963
|
| 52 |
+
},
|
| 53 |
+
{
|
| 54 |
+
"format": "parquet",
|
| 55 |
+
"path": "datapool.train.parquet",
|
| 56 |
+
"row_count": 312,
|
| 57 |
+
"row_id_sha256": "2571fe9e2e8b5e73e4b3705147ace20409c40679ff959a1f94d6103f6b6f1df6",
|
| 58 |
+
"sha256": "beb0f3ebc295f967dc61dc34628bf20b5fff93de491b6fabe610e7fb47e46210",
|
| 59 |
+
"size_bytes": 235090
|
| 60 |
+
},
|
| 61 |
+
{
|
| 62 |
+
"format": "jsonl",
|
| 63 |
+
"path": "datapool.train.sampled.jsonl",
|
| 64 |
+
"row_count": 312,
|
| 65 |
+
"row_id_sha256": "2571fe9e2e8b5e73e4b3705147ace20409c40679ff959a1f94d6103f6b6f1df6",
|
| 66 |
+
"sha256": "d2244e1dedda59d4f5d7399a6be0beac933a3d28c30e799de0dedcef3b6ff2f0",
|
| 67 |
+
"size_bytes": 629963
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"format": "parquet",
|
| 71 |
+
"path": "datapool.train.sampled.parquet",
|
| 72 |
+
"row_count": 312,
|
| 73 |
+
"row_id_sha256": "2571fe9e2e8b5e73e4b3705147ace20409c40679ff959a1f94d6103f6b6f1df6",
|
| 74 |
+
"sha256": "beb0f3ebc295f967dc61dc34628bf20b5fff93de491b6fabe610e7fb47e46210",
|
| 75 |
+
"size_bytes": 235090
|
| 76 |
+
},
|
| 77 |
+
{
|
| 78 |
+
"format": "jsonl",
|
| 79 |
+
"path": "ood_test.jsonl",
|
| 80 |
+
"row_count": 546,
|
| 81 |
+
"row_id_sha256": "03a0ce944e0db8682c754c4c8e66357ac3da9e81973a31b68afff008c7fd4758",
|
| 82 |
+
"sha256": "c2b8ccb613625b13865c06ebeb78157333679b32255f03e9372af9c37f1d3435",
|
| 83 |
+
"size_bytes": 290707
|
| 84 |
+
},
|
| 85 |
+
{
|
| 86 |
+
"format": "other",
|
| 87 |
+
"path": "prepare_data_summary.json",
|
| 88 |
+
"row_count": null,
|
| 89 |
+
"row_id_sha256": null,
|
| 90 |
+
"sha256": "e4a198882b9d64029c7dadd1f78f914245f7a4ff370177aa24d7537936e16bbd",
|
| 91 |
+
"size_bytes": 707
|
| 92 |
+
}
|
| 93 |
+
],
|
| 94 |
+
"schema_version": 2,
|
| 95 |
+
"source_artifacts_dir": "artifacts/data"
|
| 96 |
+
}
|