jackyk02/embed-nemotron-qwen3.6-35b-a3b-fp8-embeddings-256k

embed-nemotron retrieval pairs, state/action embeddings with mined hard negatives: Qwen3.6-35B-A3B-FP8, 256k context

Every distinct query of the 14 subsets of nvidia/embed-nemotron-dataset-v1 (revision f457c3e: the fine-tuning compilation behind llama-embed-nemotron-8b; the eleven ids-only subsets restored to text from their sources at pinned revisions) as a (state, action) embedding pair from Qwen/Qwen3.6-35B-A3B-FP8 with up to 7 of the release's mined hard negatives: the state is the subset's instruction and the query (Instruct: <task>\nQuery: <query>), the action the positive passage (or, for the classification subsets, the label text), and the negatives the first mined negatives in the miner's order, embedded like actions. For post-training contrastive (CLM) heads on retrieval and option choice.

split	pairs	tasks	negatives per pair	negative tensor
train	3,116,949	3,116,460	mean 5.72 (7: 2,167,502; 1-6: 949,437; 0: 10)	[3116949, 7, 2048]
val	347,560	347,503	mean 5.72 (7: 241,301; 1-6: 106,259)	[347560, 7, 2048]

subset (domain = embed-nemotron/<subset>)	task	train	val
SyntheticClassificationData	classification	938,062	104,979
PAQ	retrieval	899,627	100,373
MSMARCO	retrieval	452,573	50,366
MAmmoTH2	retrieval	285,378	31,798
FEVER	retrieval	92,037	10,255
NaturalQuestions	retrieval	90,350	9,881
GooAQ	retrieval	89,996	10,004
SQuAD	retrieval	78,597	8,744
HotpotQA	retrieval	76,503	8,491
TriviaQA	retrieval	65,994	7,352
MIRACL/<script> (18 languages)	retrieval	34,612	3,819
EmotionClassification	classification	11,695	1,321
NFCorpus	retrieval	800	95
SciFact	retrieval	725	82

MIRACL's domain carries the query's script (train: latin 12,659, arabic 4,843, cyrillic 4,075, japanese 3,055, telugu 3,042, thai 2,632, bengali 1,457, chinese 1,130, devanagari 949, korean 770), a script, not a language. One record per distinct query (exact duplicates merged: the union of their positives, then negatives), one positive each. 19,814,210 negatives in all, every one labelled mined; a negative equal to a positive is dropped, MIRACL's 100 per query are cut to the first 7, the classification subsets have 1-5 (the other labels). SyntheticClassificationData's queries already carry their instruction and get none added. No state or action is clipped (window 32,768; states average ~60 tokens, actions ~120, the longest 5,850).

Layout

train/state_embeddings.pt    float16 [3116949, 2048]       (torch.save; torch.load(..., mmap=True))
train/action_embeddings.pt   float16 [3116949, 2048]
train/neg_embeddings.pt      float16 [3116949, 7, 2048]    the mined negatives, embedded like actions
train/neg_counts.pt          int64   [3116949]             0-7: how many of the slots are filled
train/metadata.json          {"num_samples", "hidden_size", "model_name", "max_model_len", "state_head_tokens",
                              "renderer", "multimodal", "negatives_k": 7, "samples": [one dict per row]}
val/...                      the same, 347,560 rows

Row i of the tensors and samples[i] describe one pair; slots past neg_counts[i] are zero. Each sample has task_id (the SHA-256 of the bare query, without the instruction), trajectory_id (embed-nemotron:<subset>:<row>), step_idx (0), reward (1: the passage is the labelled right answer), domain, split, source_row_id, source_step_id, negatives and negative_labels (one per filled slot), and the original and clipped token counts. Vectors are L2-normalized (last-token pooling). The split is by task (int(task_id[:8], 16) % 10 == 0 -> val).

Rendering

Plain text, no chat template, clm_train.render.Renderer (the Diverse-QA pre-training store's token contract): the state reads User:\nInstruct: <task>\nQuery: <query> (the e5 form), e.g. Instruct: Given a web search query, retrieve relevant passages that answer the query / Query: who was the head coach of norway football team from 1990 to 1998; the action and the negatives are the passages' text, or the label (sarcasm against irony, slapstick). The MTEB corpora's title is not used. clm_train.inference.CLMScorer renders the same way at inference.

Use

clm-train posttrain --source hf:embed-nemotron-qwen3.6 --source hf:agentic-v4-qwen3.6 ...   # negatives join the loss (agentic_hard_negatives)
python scripts/download.py data embed-nemotron-qwen3.6 --rows 20000 --blocks 8

License

As the source dataset (NVIDIA): FEVER and MIRACL CC BY-SA 4.0, SyntheticClassificationData CC BY 4.0, every other subset under the licence of the dataset its texts come from. SyntheticClassificationData was generated with Llama 3.3 / Llama 4 models: a model trained on it and distributed may be subject to the Llama 3.3 and Llama 4 Community License Agreements. The source card states the dataset is for research and development.