InnerJev-4B-LoRA
InnerJev-4B-LoRA belongs to the InnerJev-4B family of decision models. It returns a probability distribution over user-supplied options with thinking disabled. Choice, binary/probabilistic Noul, and ordered Score tasks share the same answer-slot interface. This training optimizes decision distributions; it does not provide general chat SFT supervision.
Model and training
| Item | Setting |
|---|---|
| Original model | Qwen/Qwen3.5-4B |
| Adaptation | LoRA, rank 64 / alpha 128 / dropout 0; exported as merged weights |
| Training questions | 29,326 |
| Validation questions | 7,506 |
| Teacher | Original same-size model, two reasoning readouts averaged |
| Student | Thinking disabled; original tokenizer and ordered option labels |
| Learning rate | 5e-05 |
| Global batch | 64 |
| Seed | 43 |
| Exported step / one-epoch schedule | 276 / 459 |
| Updated parameters during training | 111,673,344 |
| Weight format | Complete BF16 safetensors; no adapter required for inference |
Only the listed language linear layers are adapted. Embeddings, the output head, normalization, and vision parameters remain frozen. Full-rank adaptation and LoRA use the same layer coverage. The architecture retains the original vision components, but InnerJev decision training and the supplied inference interface use text inputs.
The per-question loss is:
w * CE(q, p) + 0.5 * w * I(Score) * ((E[p]-E[q])/(K-1))² + kappa * KL(p0 || p).
Choice/Noul/Score weights are 1 / 1.1666667 / 0.8. Kappa is 0.2 for hard binary Noul,
and 0.05 otherwise. Gold labels are used for validation and identifying hard binary tasks;
they do not replace teacher targets. See training_recipe.json for the exact settings.
Evaluation
The associated BF16 checkpoint scores 73.9648 / 100 on the fixed 58-benchmark decision suite, with equal weight per benchmark. This is one checkpoint's score, not a multi-seed mean or a confidence bound. Validation metrics determine the checkpoint within its run. The LoRA score was measured before BF16 weight merging; a full 58-benchmark evaluation of the merged distribution has not been repeated. NF4/INT8 results may differ from the BF16 evaluation and have not been assigned this score.
Load with Transformers
The complete merged weights can be loaded without PEFT or a separate base model.
Install torch, accelerate, and transformers==5.14.1.
import torch
from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration
repo = "jylin001206/InnerJev-4B-LoRA"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
repo, dtype=torch.bfloat16, device_map="auto"
)
Loading in BF16 requires sufficient RAM or GPU memory. This example loads the original precision; the RTX 4090 launchers below use the stated deployment precision. Training code and training data are publicly available at the repository links below.
Inference
Use the decision scripts from https://github.com/amazingljy1206/InnerJev; generic free-form generation does not implement the decision scoring interface.
git clone https://github.com/amazingljy1206/InnerJev.git
cd InnerJev
pip install -e '.[quantization]'
bash scripts/infer_4b_single_4090.sh \
jylin001206/InnerJev-4B-LoRA examples/decisions.jsonl predictions.jsonl
The scripts work with either a Hugging Face repository ID or a local model directory. 4B uses BF16 on one RTX 4090; its two-GPU script runs two independent replicas. 27B uses NF4 on one RTX 4090 and INT8 by default on two RTX 4090 GPUs. Two 24-GB cards cannot hold the complete 27B BF16 model; the two-GPU script uses model sharding with quantized weights. Inputs are rejected rather than truncated when they exceed the configured token limit.
Data and reproduction
Training data: jylin001206/InnerJev-4B-Training-Data. Training code, exact layer lists, four method configs, and merge code: https://github.com/amazingljy1206/InnerJev. The tokenizer, chat template, configuration, license, and safetensors shards are included here.
License
The model retains the Apache-2.0 license of the original Qwen weights. See LICENSE and NOTICE.
Training-data source materials retain their upstream terms, as described in the dataset card.