Ma7ee7/Meet7_0.6b_Exp_Thinking

Meet7 0.6B — Experimental Thinking

A thinking trained variant of Meet7 0.6B Experimental, re-enabling Qwen3's built-in chain-of-thought reasoning at inference time.

Note: At 0.6B scale, thinking mode does not improve benchmark performance. (Except over base model which was tested with reasoning.) The model still lacks sufficient capacity to reason coherently across extended thought chains. For best results, use Meet7 Experimental without thinking mode, or Meet7 0.6B if BoolQ-style QA is your primary use case.

Benchmarks

0-shot evaluation across all four models in the Meet7 family. Scores are acc_norm.

Task	Base	Meet7	Experimental	Exp_Thinking
BoolQ	0.3798	0.5554	0.3991	0.3783
ARC Easy	0.3384	0.3952	0.3965	0.3662
ARC Challenge	0.2841	0.3285	0.3259	0.3012
HellaSwag	0.3981	0.4205	0.4265	0.4261
PIQA	0.6338	0.6583	0.6687	0.6540
Winogrande	0.5225	0.5201	0.5304	0.5241

What these measure

• BoolQ — Reading comprehension and yes/no factual grounding
• ARC Easy / Challenge — Grade-school science reasoning; Challenge is the retrieval-resistant subset
• HellaSwag — Commonsense sentence completion
• PIQA — Physical world intuition
• Winogrande — Commonsense pronoun resolution

Meet7 Model Family

Model	Strengths	Weaknesses
Meet7 0.6B	BoolQ (+17.56% vs base)	Less balanced overall
Meet7 Experimental	Best overall balance, wins 4/6 tasks	BoolQ regresses vs Meet7
Meet7 Exp_Thinking (this model)	Thinking mode enabled	(OUR) Weakest benchmark scores at 0.6B scale

Model Details

Developed by	Ma7ee7
License	Apache-2.0
Base model	Ma7ee7/Meet7_0.6b-experimental
Original base	unsloth/Qwen3-0.6B-unsloth-bnb-4bit
Training samples	1800
Thinking mode	Enabled (Qwen3 native)

Trained 2x faster with Unsloth and Hugging Face TRL.