Cognius-mini-1B-v2

A LoRA fine-tune of meta-llama/Llama-3.2-1B-Instruct, trained via supervised fine-tuning to produce structured chain-of-thought and self-critique before answering, on a small curated reasoning dataset.

Model Details

Model Description

Cognius-mini-1B-v2 is Llama-3.2-1B-Instruct fine-tuned with LoRA on 1,044 curated (question, reasoning, self-reflection, answer) examples. The training format asks the model to think through a question, critique its own reasoning, and then answer — rather than answering directly. Evaluated against the same-size untrained base model on 42 held-out questions, judged pairwise by an LLM on a 5-criterion rubric, the fine-tuned model scored 40.8% higher on average (8.68/10 vs. 6.16/10 for the base model).

• Developed by: harshil30402
• Model type: Causal decoder-only LM, LoRA adapter over Llama-3.2-1B-Instruct
• Language(s): English
• License: MIT
• Finetuned from model: meta-llama/Llama-3.2-1B-Instruct

Model Sources

• Repository: https://huggingface.co/harshil30402/Cognius-mini-1B-v2
• Training/eval code: https://github.com/lucidopus/Cognius-1B

Uses

Direct Use

Short-form question answering where a brief, structured reasoning + self-reflection pass before the final answer is useful — e.g. as a lightweight local reasoning assistant. The model was trained on short, general-knowledge questions (science, history, concepts), not on math word problems or formal logic puzzles.

Out-of-Scope Use

Not evaluated for factual reliability on specialized/technical domains, multi-turn dialogue, or long-context tasks. Not suitable for use where a wrong or fabricated factual answer could cause harm — outputs are not fact-checked or grounded.

Training Details

Training Data

• 1,044 examples, each with four fields: task (the question), reasoning (a chain-of-thought), self_reflection (a critique/refinement of that reasoning), and answer (the final answer). Topics are general open-domain knowledge questions (e.g. "How does photosynthesis work?", "What is artificial intelligence?"), not benchmark math/logic sets.
• Split into 835 train / 104 val / 105 test at the raw-example level, then reformatted into a Llama-3 chat template; after formatting, 820 train / 164 val examples were actually used for training.
• Known limitation: the held-out test split was not fully de-duplicated against training data — roughly 29% of test questions share task text with a training example (see Evaluation). Treat the evaluation numbers below as directionally correct rather than a clean generalization measurement.

Training Procedure

Supervised fine-tuning with TRL's SFTTrainer and a completion-only loss mask (DataCollatorForCompletionOnlyLM, splitting on <|end_header_id|> so loss is only computed on the assistant's response, not the prompt).

Preprocessing

• Tokenizer: Llama-3.2-1B-Instruct's tokenizer with an added <|pad|> pad token (right-padded); vocabulary resized to 128,264 (pad_to_multiple_of=8 alignment).
• Chat template wraps the question, a <think>...</think> reasoning block, and a <self_reflection>...</self_reflection> block ahead of the final answer, with a system prompt instructing the model to "think and express meta cognition" before answering.
• max_seq_length: 512 tokens.

Training Hyperparameters

Hyperparameter	Value
LoRA rank (r)	32
LoRA alpha	16
LoRA dropout	0.005
LoRA target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
LoRA bias	none
Learning rate	1e-4
LR scheduler	constant, 10% warmup
Optimizer	AdamW (torch), β=(0.9, 0.999), ε=1e-8
Effective batch size	8 (per-device 2 × gradient accumulation 4)
Epochs / steps	~1 epoch (102 steps)
Precision	fp32
Weight decay	0.0
Max grad norm	1.0
Seed	42

Speeds, Sizes, Times

Trained locally on Apple Silicon (MPS), not a cloud GPU. Train loss fell from 1.18 to 0.86 over the 102 steps.

Evaluation

Testing Data, Factors & Metrics

• Testing data: 42 held-out questions (see the data-contamination caveat above).
• Method: for each question, both the untrained base model (Llama-3.2-1B-Instruct) and this fine-tuned model generated one response. An LLM judge (openai/gpt-oss-120b, served via Groq) scored each response 0–10 against the ground-truth answer on five criteria — Accuracy, Completeness, Clarity & Coherence, Conciseness & Relevance, and Depth of Reasoning — and the five scores were averaged into one overall score per response.
• Note on judge model: the original judge planned for this project (Llama-3.3-70B) was retired from Groq's API by the time this evaluation was run; openai/gpt-oss-120b (120B open-weight model) was used instead as the closest available substitute, with an identical rubric.

Results

	Mean score (0–10)
Base model (Llama-3.2-1B-Instruct, untrained)	6.16
Cognius-mini-1B-v2 (fine-tuned)	8.68
Relative improvement	+40.8%

Summary

The fine-tuned model consistently produced shorter, more on-target answers matching the ground-truth style, versus the base model's more verbose, sometimes tangential, and occasionally erroneous (e.g. false-refusal) responses. Given the test-set overlap noted above, treat the magnitude as an upper-bound estimate rather than a precise generalization number.

How to Get Started with the Model

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_name = "meta-llama/Llama-3.2-1B-Instruct"
adapter_name = "harshil30402/Cognius-mini-1B-v2"

tokenizer = AutoTokenizer.from_pretrained(adapter_name)
base_model = AutoModelForCausalLM.from_pretrained(base_model_name)
base_model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=8, mean_resizing=False)

model = PeftModel.from_pretrained(base_model, adapter_name)
model.eval()

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "What is artificial intelligence?"}],
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Bias, Risks, and Limitations

• Evaluated on a small (42-example), partially train-overlapping test set — treat results as indicative, not a rigorous benchmark.
• Judged by a single LLM judge rather than multiple independent judges or human raters; no inter-rater agreement was measured.
• Trained on general-knowledge trivia-style questions; not evaluated on mathematical, logical, or multi-step reasoning benchmarks despite the "reasoning" framing.
• Outputs are not fact-checked; the model can still produce incorrect or fabricated answers.

Technical Specifications

Compute Infrastructure

• Hardware: Apple Silicon (MPS), local machine.
• Software: PyTorch 2.6.0, Transformers 4.50.0, PEFT 0.15.0, TRL.

Model Card Contact

harshil30402