yannelli/d1-3b-code-review-decisions

d1-3B Code Review Decisions

A decision model that reviews JavaScript and TypeScript code. Give it a snippet and the task the snippet implements, and it answers 13 typed questions with probabilities: which of 10 code review flaws the code has, whether it has any flaw, which flaw matters most, and how mergeable it is. It reads the snippet once and answers all 13 questions as branches of one forward pass, without generating text. Of the four models trained on this dataset it scores highest on every metric.

It is LiquidAI/d1-3B fine-tuned on JS/TS Code Review Decisions. On that dataset's test split it scores macro F1 0.917 and accuracy 0.940 over 15,600 decisions.

Base model	LiquidAI/d1-3B, LFM2.5-VL-3B with Liquid AI's d1 decision readout
Parameters	3.1B (bf16, 6.2 GB)
Input	A state with task, language and code, and named questions
Output	P(yes) for yes/no questions; a distribution over options for choice and score questions
Training data	JS/TS Code Review Decisions, decisions config, train split (9,600 snippets)
Test results	Macro F1 0.917 (cutoff 0.5), accuracy 0.940, ECE 0.017
License	LFM Open License v1.0

Questions

The model answers the 13 questions of the dataset's decisions config. They are the same for every snippet and are stored in the dataset's manifest.json. Each question's criteria carry the flaw definition the labels follow.

Question	Type	Asks	Answer
unclear_naming	noul	Does this code have unclear naming?	P(yes)
noise_comments	noul	Does this code have noise comments?	P(yes)
dead_code	noul	Does this code have dead code?	P(yes)
inconsistent_style	noul	Does this code have inconsistent style?	P(yes)
over_abstracted	noul	Does this code have over-abstraction?	P(yes)
too_clever	noul	Does this code have overly clever code?	P(yes)
duplicated_logic	noul	Does this code have duplicated logic?	P(yes)
error_handling_problem	noul	Does this code have an error handling problem?	P(yes)
type_safety_hole	noul	Does this code have a type safety hole?	P(yes)
unsafe_input	noul	Does this code have unsafe input handling?	P(yes)
has_flaw	noul	Does this code have at least one code review flaw?	P(yes)
primary_flaw	choice	Which code review problem matters most in this code? Choose none if the code is clean.	Distribution over none and the 10 flaws
quality	score	How mergeable is this code as written?	Distribution over 5 levels, 0 "Should not merge." to 4 "Exemplary. Merge as is."

Quality level i is quality i + 1 in the dataset's default config.

Usage

pip install torch==2.14.1 transformers==5.19.0 huggingface_hub safetensors

import json

import torch
from huggingface_hub import hf_hub_download
from transformers import AutoModel

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32
model = AutoModel.from_pretrained("yannelli/d1-3b-code-review-decisions",
                                  trust_remote_code=True, dtype=dtype).to(device)

# The 13 questions the model was trained on, from the dataset's manifest.json
manifest = hf_hub_download("yannelli/js-ts-code-review-decisions", "manifest.json", repo_type="dataset")
questions = json.load(open(manifest))["questions"]

state = {
    "task": "Define payroll bank account fields and decode a JSON account record, "
            "accepting only a nine-digit routing number and a four-digit account suffix.",
    "language": "TypeScript",
    "code": """import { z } from 'zod';

const bankAccountSchema = z.object({
    routingNumber: z.string().regex(/^\\d{9}$/),
    accountSuffix: z.string().regex(/^\\d{4}$/)
});
export type PayrollBankAccount = z.infer<typeof bankAccountSchema>;
export function decodeBankAccount(serialized: string): PayrollBankAccount {
    const account_payload = JSON.parse(serialized);
    return account_payload;
}""",
}

answers = model.system_one(state, questions)["answers"]
print(json.dumps(answers, indent=1))

The example is a test-split snippet the model did not see in training. Its gold labels are inconsistent_style (account_payload next to camelCase names) and unsafe_input (the JSON.parse result is returned without the schema defined above it). This model's answers, from the code above on CPU:

Question	Gold	This model
unclear_naming	no	P(yes) 0.00
noise_comments	no	P(yes) 0.00
dead_code	no	P(yes) 0.03
inconsistent_style	yes	P(yes) 0.99
over_abstracted	no	P(yes) 0.00
too_clever	no	P(yes) 0.00
duplicated_logic	no	P(yes) 0.00
error_handling_problem	no	P(yes) 0.01
type_safety_hole	no	P(yes) 0.01
unsafe_input	yes	P(yes) 0.99
has_flaw	yes	P(yes) 1.00
primary_flaw	unsafe_input	unsafe_input (0.98)
quality	level 0	level 1 (0.51); expected level 0.62

system_one returns the System One answer shape. A noul answer is P(yes). A choice answer has the top choice, its confidence, and probabilities over all options. A score answer has score (the expected level), confidence, probabilities over levels, and the legend. model.system_one_batch answers several snippets in one call. The quality answer above, rounded:

{
 "type": "score",
 "score": 0.623,
 "confidence": 0.509,
 "probabilities": {
  "0": 0.435,
  "1": 0.509,
  "2": 0.055,
  "3": 0.001,
  "4": 0.0
 },
 "legend": {
  "0": "Should not merge.",
  "1": "Needs significant changes before it can merge.",
  "2": "Mergeable with review comments.",
  "3": "Mergeable with minor comments.",
  "4": "Exemplary. Merge as is."
 }
}

On Apple Silicon set device = "mps" (not benchmarked). On CPU the model runs in float32. For faster CPU inference, or to run under llama.cpp or Ollama, use the GGUF builds.

The model accepts other questions in the same schema, but it was trained and evaluated only on these 13.

Results

Results cover the test split of the decisions config: 300 tasks, 1,200 snippets, and 15,600 decisions, scored with the dataset's score_decisions.py. Each snippet was one request with all 13 questions. The four fine-tuned models used the same data and scorer.

Model	Params	Macro F1 @ 0.5	Accuracy	Log loss	Brier	ECE	Clean judged clean
d1-3B (this model)	3.1B	0.917	0.940	0.172	0.072	0.017	0.950
d1-omni-600M	587M	0.771	0.882	0.321	0.127	0.028	0.827
GLiNER2.5-Decide	486M	0.687	0.863	0.398	0.149	0.039	0.683
NanoJev	596M	0.452	0.831	0.459	0.173	0.038	0.000
d1-3B base model, zero-shot	3.1B	0.220	0.726	0.679	0.249	0.142	0.953
TF-IDF + logistic regression (dataset baseline)	-	0.502	0.830	0.449	0.172	0.030	0.307

Higher is better for macro F1, accuracy, and clean judged clean; lower is better for log loss, Brier, and ECE. Clean judged clean is the share of the 300 clean test snippets with P(has_flaw) below 0.5. The dataset card's Metrics section defines the other metrics.

95% intervals for this model, from 500 bootstrap resamples of test tasks: macro F1 0.904 to 0.928, accuracy 0.935 to 0.946, log loss 0.158 to 0.189, Brier 0.067 to 0.078. By question type its accuracy is 0.975 on the 11 nouls, 0.812 on primary_flaw, and 0.684 on quality.

Per question

Question	Type	Test positives	Accuracy	F1 @ 0.5	Log loss	ECE
unclear_naming	noul	139	0.992	0.964	0.030	0.006
noise_comments	noul	178	0.999	0.997	0.002	0.001
dead_code	noul	223	0.961	0.886	0.154	0.013
inconsistent_style	noul	147	0.977	0.903	0.079	0.009
over_abstracted	noul	145	0.988	0.952	0.038	0.009
too_clever	noul	164	0.981	0.929	0.045	0.010
duplicated_logic	noul	171	0.963	0.867	0.112	0.020
error_handling_problem	noul	140	0.988	0.946	0.037	0.009
type_safety_hole	noul	148	0.973	0.879	0.106	0.015
unsafe_input	noul	165	0.962	0.845	0.142	0.021
has_flaw	noul	900	0.946	0.963	0.146	0.024
primary_flaw	choice		0.812		0.591	0.035
quality	score		0.684		0.760	0.051

Calibration

Answers are the model's raw softmax (test ECE 0.017). temperatures.json holds one temperature per question, fitted on the validation split by log loss. Replacing each answer's distribution p with softmax(log p / T) moves test ECE from 0.017 to 0.015. It never changes the top answer or the order of P(yes) values, so F1 and accuracy stay the same.

Speed

Median wall-clock time for one snippet with all 13 questions, batch size 1. The PyTorch CPU row covers 52 test snippets, the llama.cpp CPU row the first 20, and the L40S row all 1,200.

Hardware	Runtime	p50 per snippet
One H100 SXM	PyTorch, bf16	30.5 ms
16 Neoverse V3 CPU cores	PyTorch, fp32	8.9 s
16 Neoverse V3 CPU cores	llama.cpp llama-server, Q8_0 GGUF	6.2 s
One L40S	llama.cpp llama-server, Q8_0 GGUF	221 ms

Intended use

Flag code review problems in JavaScript and TypeScript while code is written or before it reaches review: in an editor, a pre-commit hook, or CI, as hints for a human reviewer. The probabilities let you pick a cutoff per flag and trade false positives against misses.

Training

• Data: the train split of the decisions config, 2,400 tasks and 9,600 snippets (124,800 decisions). The state is the task description, language and code. The training files match the SHA-256 hashes in the dataset's published manifest.json.
• The dataset keeps the 4 variants of a task in one split, so no test task was seen in training.

Setting	Value
Method	LoRA (r 32, alpha 64, dropout 0.05) on the language model's attention, convolution and MLP projections, merged into the weights after training
Objective	Cross entropy over each question's answer tokens; the snippet is the trunk of d1's packed tree and the 13 questions are its branches
Schedule	1 epoch, 1,200 steps of 8 snippets (104 decisions)
Optimizer	AdamW, learning rate 2e-4, 5% warmup, linear decay
Precision	bf16
Hardware	One H100 SXM, 36 minutes
Validation loss	0.207 at the end of training (300 validation snippets)

training/ holds the scripts that trained and scored this model (liquid_d1.py, common.py, calibrate.py). To rerun it:

pip install peft==0.21.2 pyarrow numpy
hf download yannelli/js-ts-code-review-decisions --repo-type dataset --local-dir js-ts-code-review-decisions
export DECISIONS_DATA=js-ts-code-review-decisions
python training/liquid_d1.py train --model LiquidAI/d1-3B --ckpt ckpt --epochs 1
python training/liquid_d1.py predict --ckpt ckpt --split test --out test-answers.jsonl
python js-ts-code-review-decisions/score_decisions.py test-answers.jsonl --split test

Evaluation files

eval/test-answers.jsonl holds this model's answers to all 1,200 test snippets in the scorer's format, and eval/score.txt is the scorer's output for it. To rescore:

python js-ts-code-review-decisions/score_decisions.py eval/test-answers.jsonl --split test

Limitations

• Gold labels come from blind model review and a model audit that accepted 98.3% of findings, not from human reviewers. Scores measure agreement with those labels.
• The snippets are synthetic, 5 to 80 lines long, and 25% of them are clean, so false positive rates on real repositories will differ.
• Training covered only JavaScript, JSX, TypeScript and TSX snippets, each with a one-sentence English task; pass one with every snippet.
• Weakest flags by test F1: unsafe_input (0.84), duplicated_logic (0.87), type_safety_hole (0.88). quality is the hardest question (accuracy 0.684); 80% of its wrong answers are one level off.
• unsafe_input follows the dataset's definition (external data used unchecked); it is not a vulnerability scanner.

Other models trained on this dataset

Model	Base	Params	Macro F1 @ 0.5	Accuracy
d1-3B (this model)	LFM2.5-VL-3B with Liquid AI's d1 decision readout	3.1B	0.917	0.940
d1-omni-600M	LFM2.5 encoder with a decision head	587M	0.771	0.882
GLiNER2.5-Decide	GLiNER2 DeBERTa-v3 encoder with a classification head	486M	0.687	0.863
NanoJev	Qwen3-0.6B with a choice head	596M	0.452	0.831

yannelli/code-review-decisions-GGUF has GGUF builds of the two Liquid models for llama.cpp and Ollama.

License

This model is a derivative of LiquidAI/d1-3B and is released under the base model's LFM Open License v1.0, included as LICENSE. The license allows commercial use only for entities with less than $10M in annual revenue; read LICENSE for the full terms.

The training data, JS/TS Code Review Decisions, is CC BY 4.0.

Citation

@misc{yannelli2026d13bcodereview,
  title        = {d1-3B Code Review Decisions},
  author       = {Ryan Yannelli},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/yannelli/d1-3b-code-review-decisions}}
}

@misc{yannelli2026jstscodereviewdecisions,
  title        = {JS/TS Code Review Decisions},
  author       = {Ryan Yannelli},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/yannelli/js-ts-code-review-decisions}}
}

Maintainer

Ryan Yannelli | 10/12/2026 | GitHub: @yannelli | Hugging Face: yannelli