engram-retrieval-v0

Cross-encoder used by the Engram knowledge-object (KO) retrieval pipeline. Given a user query and a candidate KO body, scores how well the KO answers the query. Used as the first-pass ranker before the more expensive engram-xenc-v0 re-ranker.

Engram context

• Codebase consumer: sidecar/gating_cross_encoder.py and the tools/session7_finetune.py evaluation harness.
• Fetch via: scripts/fetch_models.sh in the Engram repo (idempotent post-clone install step).
• Reproduce locally: python tools/session7_finetune.py (~5 minutes on M4 Max). Inputs are data/session7_test_set.json (244 labels, 243 used after reject-exclusion) + wiki/ (KO bodies resolved via WikiFileStore at train time). Deterministic seed (SEED=7), stratified 80/20 split.
• Eval method: stratified by (label, domain). AUC reported to stdout by the training script; ROC sweep at 200 thresholds.
• Drift caveat: training reads ko.content from wiki/ at train time (not snapshotted into the test set). All 243 ko_ids resolve in the current wiki/ as of 2026-06-25; future wiki edits could silently change re-train results. Snapshotting is tracked as a follow-up in docs/decisions/2026-06-25-model-distribution.md.

License

Apache 2.0. Derived from cross-encoder/ms-marco-MiniLM-L6-v2 (Apache 2.0).

CrossEncoder based on cross-encoder/ms-marco-MiniLM-L6-v2

This is a Cross Encoder model finetuned from cross-encoder/ms-marco-MiniLM-L6-v2 using the sentence-transformers library. It computes scores for pairs of texts, which can be used for text reranking and semantic search.

Model Details

Model Description

• Model Type: Cross Encoder
• Base model: cross-encoder/ms-marco-MiniLM-L6-v2
• Maximum Sequence Length: 512 tokens
• Number of Output Labels: 1 label
• Supported Modality: Text

Model Sources

• Documentation: Sentence Transformers Documentation
• Documentation: Cross Encoder Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Cross Encoders on Hugging Face

Full Model Architecture

CrossEncoder(
  (0): Transformer({'transformer_task': 'sequence-classification', 'modality_config': {'text': {'method': 'forward', 'method_output_name': 'logits'}}, 'module_output_name': 'scores', 'architecture': 'BertForSequenceClassification'})
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import CrossEncoder

# Download from the 🤗 Hub
model = CrossEncoder("cross_encoder_model_id")
# Get scores for pairs of inputs
pairs = [
    ['How might AI automation affect the way developers work?', '> **Type:** constraint — A limitation, dependency, or hard requirement.  \n> **Domain:** engineering  \n> **Scope:** team  \n> **Confidence:** 60%  \n> **Created:** 2026-04-26 23:23 UTC by dave\n\nStart the session by reading init.sh . Summarizing four common failure modes and solutions in long-running AI agents.'],
    ['Is there a risk that our technical depth declines if we lean too hard on AI assistants?', '> **Type:** constraint — A limitation, dependency, or hard requirement.  \n> **Domain:** engineering  \n> **Scope:** team  \n> **Confidence:** 70%  \n> **Created:** 2026-04-26 23:23 UTC by dave\n\nSome engineers worry that relying on AI will lead to a loss of deeper technical skills.'],
    ['What are the key dates and deliverables we need to hit during our quarterly financial close?', '> **Type:** skill — A procedure, how-to, or repeatable technical process.  \n> **Domain:** finance  \n> **Scope:** restricted  \n> **Confidence:** 90%  \n> **Created:** 2026-04-24 22:27 UTC by erin\n\nQuarterly close sequence: revenue recognition review by day 3, AR aging reconciliation by day 5, accrual entries locked day 7, management pack draft day 10, board-ready numbers day 14.'],
    ['How do we test whether our coding agents are robust enough for production?', '> **Type:** question — An open question that has not yet been answered.  \n> **Domain:** engineering  \n> **Scope:** org  \n> **Confidence:** 70%  \n> **Created:** 2026-04-26 23:23 UTC by dave\n\nMost notably, it’s still unclear whether a single, general-purpose coding agent performs best across contexts, or if better performance can be achieved through a multi-agent architecture.'],
    ['What are the main concerns engineers have about AI replacing their skills?', '> **Type:** policy — An organisational rule or standard that must be followed.  \n> **Domain:** engineering  \n> **Scope:** org  \n> **Confidence:** 90%  \n> **Created:** 2026-04-26 23:23 UTC by dave\n\nThe research highlights the need for guidelines and frameworks to support the adoption of AI in software development.'],
]
scores = model.predict(pairs)
print(scores)
# [-6.3638  2.7978  0.6201 -0.0405 -2.6479]

# Or rank different texts based on similarity to a single text
ranks = model.rank(
    'How might AI automation affect the way developers work?',
    [
        '> **Type:** constraint — A limitation, dependency, or hard requirement.  \n> **Domain:** engineering  \n> **Scope:** team  \n> **Confidence:** 60%  \n> **Created:** 2026-04-26 23:23 UTC by dave\n\nStart the session by reading init.sh . Summarizing four common failure modes and solutions in long-running AI agents.',
        '> **Type:** constraint — A limitation, dependency, or hard requirement.  \n> **Domain:** engineering  \n> **Scope:** team  \n> **Confidence:** 70%  \n> **Created:** 2026-04-26 23:23 UTC by dave\n\nSome engineers worry that relying on AI will lead to a loss of deeper technical skills.',
        '> **Type:** skill — A procedure, how-to, or repeatable technical process.  \n> **Domain:** finance  \n> **Scope:** restricted  \n> **Confidence:** 90%  \n> **Created:** 2026-04-24 22:27 UTC by erin\n\nQuarterly close sequence: revenue recognition review by day 3, AR aging reconciliation by day 5, accrual entries locked day 7, management pack draft day 10, board-ready numbers day 14.',
        '> **Type:** question — An open question that has not yet been answered.  \n> **Domain:** engineering  \n> **Scope:** org  \n> **Confidence:** 70%  \n> **Created:** 2026-04-26 23:23 UTC by dave\n\nMost notably, it’s still unclear whether a single, general-purpose coding agent performs best across contexts, or if better performance can be achieved through a multi-agent architecture.',
        '> **Type:** policy — An organisational rule or standard that must be followed.  \n> **Domain:** engineering  \n> **Scope:** org  \n> **Confidence:** 90%  \n> **Created:** 2026-04-26 23:23 UTC by dave\n\nThe research highlights the need for guidelines and frameworks to support the adoption of AI in software development.',
    ]
)
# [{'corpus_id': ..., 'score': ...}, {'corpus_id': ..., 'score': ...}, ...]

Training Details

Training Dataset

Unnamed Dataset

• Size: 194 training samples
• Columns: sentence_0, sentence_1, and label
• Approximate statistics based on the first 194 samples:

  	sentence_0	sentence_1	label
  type	string	string	float
  details	min: 9 tokensmean: 17.73 tokensmax: 34 tokens	min: 65 tokensmean: 117.34 tokensmax: 198 tokens	min: 0.0mean: 0.61max: 1.0

• Samples:

  sentence_0	sentence_1	label
  How might AI automation affect the way developers work?	> Type: constraint — A limitation, dependency, or hard requirement. > Domain: engineering > Scope: team > Confidence: 60% > Created: 2026-04-26 23:23 UTC by daveStart the session by reading init.sh . Summarizing four common failure modes and solutions in long-running AI agents.	0.0
  Is there a risk that our technical depth declines if we lean too hard on AI assistants?	> Type: constraint — A limitation, dependency, or hard requirement. > Domain: engineering > Scope: team > Confidence: 70% > Created: 2026-04-26 23:23 UTC by daveSome engineers worry that relying on AI will lead to a loss of deeper technical skills.	1.0
  What are the key dates and deliverables we need to hit during our quarterly financial close?	> Type: skill — A procedure, how-to, or repeatable technical process. > Domain: finance > Scope: restricted > Confidence: 90% > Created: 2026-04-24 22:27 UTC by erinQuarterly close sequence: revenue recognition review by day 3, AR aging reconciliation by day 5, accrual entries locked day 7, management pack draft day 10, board-ready numbers day 14.	0.0

• Loss: BinaryCrossEntropyLoss with these parameters:

  {
      "activation_fn": "torch.nn.modules.linear.Identity",
      "pos_weight": null
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 16
• num_train_epochs: 1
• per_device_eval_batch_size: 16

All Hyperparameters

• per_device_train_batch_size: 16
• num_train_epochs: 1
• max_steps: -1
• learning_rate: 5e-05
• lr_scheduler_type: linear
• lr_scheduler_kwargs: None
• warmup_steps: 0
• optim: adamw_torch_fused
• optim_args: None
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• optim_target_modules: None
• gradient_accumulation_steps: 1
• average_tokens_across_devices: True
• max_grad_norm: 1
• label_smoothing_factor: 0.0
• bf16: False
• fp16: False
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• use_liger_kernel: False
• liger_kernel_config: None
• use_cache: False
• neftune_noise_alpha: None
• torch_empty_cache_steps: None
• auto_find_batch_size: False
• log_on_each_node: True
• logging_nan_inf_filter: True
• include_num_input_tokens_seen: no
• log_level: passive
• log_level_replica: warning
• disable_tqdm: False
• project: huggingface
• trackio_space_id: None
• trackio_bucket_id: None
• trackio_static_space_id: None
• per_device_eval_batch_size: 16
• prediction_loss_only: True
• eval_on_start: False
• eval_do_concat_batches: True
• eval_use_gather_object: False
• eval_accumulation_steps: None
• include_for_metrics: []
• batch_eval_metrics: False
• save_only_model: False
• save_on_each_node: False
• enable_jit_checkpoint: False
• push_to_hub: False
• hub_private_repo: None
• hub_model_id: None
• hub_strategy: every_save
• hub_always_push: False
• hub_revision: None
• load_best_model_at_end: False
• ignore_data_skip: False
• restore_callback_states_from_checkpoint: False
• full_determinism: False
• seed: 42
• data_seed: None
• use_cpu: False
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• parallelism_config: None
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• dataloader_prefetch_factor: None
• remove_unused_columns: True
• label_names: None
• train_sampling_strategy: random
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• ddp_static_graph: None
• ddp_backend: None
• ddp_timeout: 1800
• fsdp: []
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• deepspeed: None
• debug: []
• skip_memory_metrics: True
• do_predict: False
• resume_from_checkpoint: None
• warmup_ratio: None
• local_rank: -1
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Time

• Training: 6.0 seconds

Framework Versions

• Python: 3.11.14
• Sentence Transformers: 5.4.1
• Transformers: 5.6.0
• PyTorch: 2.11.0
• Accelerate: 1.13.0
• Datasets: 4.8.5
• Tokenizers: 0.22.2

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}