oneiroNER

Named entity recognition for conversational text. Extracts 6 sync entity types from multi-turn conversations, optimized for low-latency inference across 75 languages.

Status

Component	State	Notes
Data pipeline	Complete	5.0M train / 266K val records, 85K silver-en, 4.2K silver-ml (46 langs)
Model harness	Validated	Local smoke + Modal GPU + W&B e2e passed (report)
Autoresearch	Validated	launch.py orchestration tested, 7 bugs fixed during validation
Dataset loading	Optimized	Lazy sidecar index + on-demand tokenization (~3 min startup vs ~20 min eager)
Training	Ready	H100 pilot passed: 18-22 steps/sec, 12.8GB VRAM, buckets match targets

Entity Types

6 sync types extracted as BIO spans (43 labels total):

Type	Subtypes
PERSON	—
PLACE	—
ORG	—
DATE	Day, Week, Month, Season, Year, Decade, Relative, Range
EVENT	Life, General
RELATIONSHIP_REF	Family, Romantic, Friend, Professional, Acquaintance

3 async types (EMOTION, GOAL, ACTIVITY) are claim-level Dreamer outputs, not span extraction.

Architecture

• Backbone: mGTE 306M (Alibaba-NLP/gte-multilingual-mlm-base) — 12 layers, 768 hidden, 8K context, 75 languages
• Head: Dropout + Linear(768, 43) token classifier
• Training: HF Trainer with MixtureSampler (75% gold / 20% silver-en / 5% silver-ml)
• Eval: Exact char-span (type, start, end) F1 — token-level is debug only
• Inference: BF16 train → CPU INT8 + GPU FP16

Directory Structure

model/              NER model, training, eval (7 files, ~1100 LOC)
research/           Autoresearch harness (7 files, ~770 LOC)
configs/            Schema, type mappings, dataset inventory, training mix
scripts/            Data pipeline: converters, annotation, validation
  lib/              Shared modules (dedup, windower, span tools, schema)
  task8/            JP-RP + ChatHaruhi batch annotation pipeline
  task9/            Silver data QA/validation pipeline
  task9_5/          Silver quality & coverage upgrade
docs/               Design docs, task plans, research analysis
tests/              Test suite
data/               Gitignored — managed locally + HF LFS

Model (model/)

File	Purpose
config.py	43 BIO labels, type normalization, hyperparams, source tiers
ner_model.py	NerModel: AutoModel encoder + dropout + linear head
ner_dataset.py	Lazy sidecar index, on-demand tokenization, BIO alignment, collate
train.py	NerTrainer with token accounting, TokenBudgetCallback, AutoresearchCallback
eval.py	Char-span F1, token-level callback, REL_REF hard-neg analysis, run_full_eval()
verify_split.py	Train/val split leakage audit

Autoresearch (research/)

Autonomous hyperparameter search adapted from karpathy/autoresearch. Runs experiments on Modal H100, tracks via W&B, uses git worktree isolation.

• Editable surface: research/train.py (EXPERIMENT dict only — nothing else)
• Composite score: weighted REL_REF F0.5 (0.40) + macro F1 (0.25) + hard-neg precision (0.20) + multilingual (0.10) + latency (0.05)
• Budget: 3.2M tokens/experiment, 60 min Modal timeout, ~$0.85/run on H100
• Stopping: no improvement for 5 consecutive experiments

See research/README.md for the full protocol.

Running

# Preparation (once)
python3 -m research.prepare

# Manual single experiment
python3 -m research.launch

# Autonomous loop (requires OpenProse)
prose run research/autoresearch.prose

Data Pipeline

Sources

Tier	Sources	Weight
Gold	Existing labeled data (~4.9M train / ~262K val)	1.0
T1	mentalchat, therapy, personachat, prosocial	0.9
T2	reddit_confessions	0.7
T3	pippa, opencharacter, roleplay_hieu, synthetic_persona_chat	0.5
Multilingual	46 languages, synthetic seed (4,960 convos)	0.8

Pipeline

# Full conversion + dedup + split
python3 scripts/convert_all.py

# Silver annotation (Task 9)
python3 scripts/task9/annotate.py

# Batch annotation (Task 8) — current pipeline
python3 scripts/task8/annotate_batch.py

Note: scripts/task8/label_roleplay.py is legacy (initial provider="spark" pipeline). The current Task 8 pipeline uses annotate_batch.py with GPT-5.4 and DeepSeek V3.2.

Configs

File	Purpose
schema.json	JSONL record format (JSON Schema draft 2020-12)
type_mapping_train.json	Source labels → canonical types for training
type_mapping_eval.json	Source labels → canonical types for eval
dataset_inventory.json	All ingested datasets with metadata
training_mix.json	Dataset mixing configuration
zero_shot_holdout_types.json	Types excluded from training

Requirements

pip install -r requirements.txt

Key dependencies: torch, transformers, accelerate, wandb, datasets, huggingface_hub

Compute

• Training: Modal (H100 $3.95/hr, B200 $6.25/hr)
• Tracking: Weights & Biases (oneiron-dev/ner-sft)
• Inference: Vast.ai Serverless (3-region: US/EU/Asia)