RenShiPDev/LFM2.5-1.2B-Thinking-LeetCode-QLoRA

LFM2.5-1.2B-Thinking + LeetCode (QLoRA)

Early-stage reasoning fine-tune (500 steps). A LoRA adapter for LiquidAI/LFM2.5-1.2B-Thinking, trained on greengerong/leetcode to study whether algorithmic/code data transfers into better logical reasoning on a small (1.2B) model.

This is not a production model and not a "better everywhere" upgrade. It is a transparent research checkpoint with a clear trade-off: improved reasoning (ARC, MMLU) at the cost of some world-knowledge recall (CommonsenseQA, OpenBookQA).

---

🧪 Training Details

Item	Value
Base model	LiquidAI/LFM2.5-1.2B-Thinking (1.17B params)
Dataset	greengerong/leetcode (code problems + solutions)
Method	QLoRA (4-bit NF4 quantization)
LoRA rank (r)	16
LoRA alpha	32
LoRA dropout	0.0
LoRA init	LoftQ
Optimizer	AdamW 8-bit
Learning rate	2e-4 (linear schedule)
Batch size	1 per device (gradient accumulation × 4 → effective batch 4)
Max steps	500
Warmup steps	3
Context length	8192 tokens
Precision	bfloat16
Seed	3407
Packing	enabled
Gradient checkpointing	enabled
Hardware	Single NVIDIA GeForce RTX 3050, 8 GB VRAM
Training time	~3 hours (wall-clock)
Frameworks	Unsloth, PEFT 0.18.1, TRL 0.23.1, Transformers 5.5.0, PyTorch 2.10.0+cu130

Training dynamics

• Train loss: 2.09 → 0.80; final eval loss 0.83.
• ~15.95M input tokens seen over 500 steps.

Note. batch_size = 1 (with ×4 accumulation) is a low-resource setting that introduces gradient noise; the results below reflect that. This checkpoint is an early exploration, not a fully converged model.

---

📊 Benchmark Results (LoRA vs. base model)

Benchmark	Base	LoRA	Δ
ARC Easy	52.44%	64.31%	+11.87
MMLU	25.89%	31.34%	+5.45
HumanEval	19.51%	21.34%	+1.83
ARC Challenge	32.08%	33.11%	+1.02
Social IQa	41.40%	42.02%	+0.61
PIQA	69.31%	69.75%	+0.44
BoolQ	71.28%	71.68%	+0.40
Winogrande	54.62%	54.93%	+0.32
HellaSwag	40.87%	40.52%	-0.35
OpenBookQA	25.60%	22.40%	-3.20
CommonsenseQA	34.56%	27.60%	-6.96

Interpretation

• Positive transfer. Large gains on ARC-Easy (+11.87 pp) and MMLU (+5.45 pp) suggest the model internalized stronger reasoning patterns from the code data — reasoning generalized beyond code.
• Trade-off / catastrophic forgetting. Drops on CommonsenseQA (-6.96 pp) and OpenBookQA (-3.20 pp) are a classic symptom of a small model overwriting world knowledge with algorithmic patterns.

---

🚨 Benchmark Transparency (data-leakage warning)

The training dataset greengerong/leetcode contains coding problems. Benchmarks such as HumanEval or MBPP may contain problems that are similar or identical to those in the training set.

Therefore the +1.83% on HumanEval should be read as task adaptation, not as a fair zero-shot generalization score against models that never saw this data. For an honest assessment, evaluate this adapter on completely unseen, custom tests.

---

⚠️ Limitations

• Not for fact/knowledge retrieval. The adapter is measurably worse than the base model on open-book and commonsense tasks.
• Not a general code assistant. Only a modest HumanEval gain; the base model is not code-specialized.
• Early checkpoint. 500 steps is a small budget; results are exploratory.
• Data-leakage caveat applies to any code benchmark (see above).

---