Key Capabilities & Highlights
- Unified Text + Vision Decision Engine: Handles multi-question text states (
~60 ms for 3 questions on one GPU) and multimodal visual reasoning (KIE, InfographicVQA, CV-Bench, Winoground, CharXiv, RealWorldQA) within the same architecture.
- 8B Total / 4B Effective Decoder Architecture (
7.94B Parameters): Combines a 3.89B core transformer decoder with 2.90B Per-Layer Embeddings (PLE), a 0.67B token embedding table, and native multimodal vision (0.17B) + audio (0.31B) encoders (14.79 GB in bfloat16).
- Block-Bidirectional Readout (
noncausal + slot_lm): Full-attention layers use a block-bidirectional mask where the shared state prefix is causal and each question/option suffix attends bidirectionally within its own block, paired with a 256-slot LM head readout (A–Z and two-letter slots).
- Calibrated Probability Distributions: Trained with a joint Cross-Entropy + Brier calibration objective (
λ = 1.0) and 1:1 soft-teacher + hard-target probability supervision, achieving 3.62% ECE on text and 8.90% ECE on multimodal vision decisions.
- Strict Benchmark Decontamination: Zero overlap with the Jev Decision Index (
0.3 / 0.4 text and vision-0.3 suites), enforced via exact dataset exclusion and SHA-256 normalized text deduplication.
Model Details
Table with columns: Property, Specification| Property | Specification |
|---|
| Model Name | Nischay (Ajeet-001/Nischay) |
| Total Parameters | 7,941,101,802 (7.94B / ~8B total) |
| Parameter Breakdown | 3.89B Core Transformer Decoder (E4B effective compute) + 2.90B Per-Layer Embeddings (PLE) + 0.67B Token Embeddings + 0.17B Vision Tower + 0.31B Audio Tower |
| Inputs | Text (structured JSON / documents / dialogues) and Images (1–4 images per request, up to 1.6 MP per image) |
| Attention Mechanism | block_bidirectional (noncausal) on full-attention layers; shared prefix cached across all questions |
| Readout Mechanism | slot_lm (256 single-token option slots, upcast float32 readout head over bfloat16 backbone) |
| Calibration | calibration.json included (T = 0.70 text / T = 1.0 multimodal) |
| Precision | bfloat16 weights (model.safetensors, 14.79 GB), float32 output head |
How to Use
1. One-Command Environment Setup (setup.sh)
setup.sh installs the pinned serving runtime (OneJev @ 81ce62f1 + serve.patch) and the official decision-index evaluation kit:
hf download Ajeet-001/Nischay --local-dir Nischay && bash Nischay/setup.sh
2. Python Quickstart (DecisionEngine)
from qev.engine import DecisionEnginefrom qev.schema import SystemOneRequest engine = DecisionEngine("Ajeet-001/Nischay", device="cuda:0") questions = { "refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"}, "team": { "type": "choice", "instructions": "Which team should handle this?", "criteria": { "billing": "Charges, refunds, invoices", "technical": "App or site faults", "fraud": "Suspected unauthorised use", }, }, "urgency": { "type": "score", "instructions": "How urgent is this?", "criteria": ["Can wait", "Today", "Blocking the customer now"], },} resp, meta = engine.decide(SystemOneRequest( state="I was charged twice this month, please refund one of them.", questions=questions,))print(resp.answers)
Table with columns: Question type, criteria Format, Returned FieldsQuestion type | criteria Format | Returned Fields |
|---|
noul (yes / no / abstention) | optional | noul: P(yes) |
choice (named options) | {key: description} | choice, probabilities |
score (ordered ordinal levels) | [lowest, ..., highest] | score, probabilities |
3. Running on the Jev Decision Index (engine.py)
This repository ships engine.py (DecisionModel), which supports both the Text and Vision suites of the Jev Decision Index out of the box:
# Text Decision Index Suite (0.3 / 0.4)PYTHONPATH=Nischay:OneJev python -m decision_index pipeline \ --engine engine:DecisionModel \ --option model=Ajeet-001/Nischay \ --out runs/nischay-text # Vision Decision Index Suite (0.3)PYTHONPATH=Nischay:OneJev python -m decision_index vision pipeline \ --dir work/vision \ --engine engine:DecisionModel \ --option model=Ajeet-001/Nischay \ --out runs/nischay-vision
Requests longer than the model's capacity limit raise Unsupported (never truncated, no option filtering, no per-benchmark prompt tuning).
Complete evaluation artifacts (results.jsonl.gz and scores.json) are published in Ajeet-001/Nischay-decision-index-results.
1. Overall Comparison with Sub-15B & Compact Decision Models
Table with columns: Model, Parameters, Modalities, Public Vision Index 0.3, Public Text Index 0.3 / 0.4| Model | Parameters | Modalities | Public Vision Index 0.3 | Public Text Index 0.3 / 0.4 |
|---|
elephant-1 | 9B | Text + Vision | — | 62.65 |
JPT-9B | 9B | Text-only | — | 46.96 |
Nischay (Ajeet-001/Nischay) | 8B (7.94B total / 3.89B E4B decoder) | Text + Vision | 44.96 (12,210 / 12,210 rows) | 37.94 (53.24% raw) |
JPT-0.8B | 0.8B | Text-only | — | 19.37 |
Lavoir | — | Text-only | — | 8.85 |
Verdict | — | Text-only | — | 1.86 |
2. Text Decision Index 0.3 — Area-Level Comparison (8B–9B Peer Class)
Table with columns: Area, Benchmarks (n), Nischay (8B / E4B), JPT-9B (9B), JPT-0.8B (0.8B)| Area | Benchmarks (n) | Nischay (8B / E4B) | JPT-9B (9B) | JPT-0.8B (0.8B) |
|---|
| Tools & Automation | 5 | 60.40 (63.68% raw) | 67.02 | 36.23 |
| Language Understanding | 9 | 41.76 (60.37% raw) | 57.68 | 25.17 |
| Retrieval & Classification | 6 | 41.73 (55.56% raw) | 44.59 | 16.40 |
| Arts & Human Taste | 6 | 29.19 (51.45% raw) | 28.78 | 8.44 |
| Knowledge & Reasoning | 11 | 18.70 (37.61% raw) | 30.92 | 8.16 |
| Overall Public Text Index | 37 | 37.94 (53.24% raw) | 46.96 | 19.37 |
3. Vision Decision Index 0.3 — Summary (12,210 / 12,210 Rows, complete: true)
Nischay is the first sub-10B multimodal decision engine evaluated on the complete 11-benchmark Official Vision Decision Index 0.3 (12,210 / 12,210 rows answered, 0 errors), achieving an overall Public Vision Index of 44.96, led by KIE (81.70 skill / 88.88% accuracy), InfographicVQA (74.32 skill / 80.74% accuracy), CV-Bench (60.50 skill / 76.05% accuracy), Winoground (60.00 skill / 80.00% accuracy), and CharXiv (49.12 skill / 61.84% accuracy), while outperforming the 27B Solomon v1.1 vision decision model on R-Bench-M (21.31 vs 18.24) and MMMU-Pro Vision (10.96 vs -2.56).
4. Held-Out Calibration & Decision Type Breakdown
Table with columns: Evaluation Split, Overall Accuracy (↑), noul (Binary / Abstention) (↑), choice (Multi-Option) (↑), score (Ordinal) (↑), NLL (↓), ECE (↓)| Evaluation Split | Overall Accuracy (↑) | noul (Binary / Abstention) (↑) | choice (Multi-Option) (↑) | score (Ordinal) (↑) | NLL (↓) | ECE (↓) |
|---|
Text Held-Out (n=200) | 85.00% | 93.20% | 78.67% | 81.82% | 0.4728 | 3.62% |
Vision Held-Out (n=150) | 56.67% | 69.57% | 53.16% | 44.00% | 1.1055 | 8.90% |
Repository Contents
Every file required to load, serve, and benchmark Nischay is included in this repository:
model.safetensors, config.json, generation_config.json: Full merged bfloat16 model weights (7.94B total parameters, 14.79 GB) and architecture configuration (with modeling_overrides for block_bidirectional full attention).
tokenizer.json, tokenizer_config.json, processor_config.json, chat_template.jinja: Multimodal processor and tokenizer configuration.
readout.json, calibration.json: Readout specification (noncausal + slot_lm, 256 slots) and temperature calibration.
engine.py: Ready-to-run DecisionModel(Engine) adapter for the Jev Decision Index (decision_index text and vision pipelines).
serve.patch, setup.sh, constraints.txt: Reproducible one-command installation of the OneJev runtime and decision-index kit.