A <thinking> token (id 151665) is added to the vocabulary and inserted into the training sequences at positions chosen by the model's own token confidence, so the model learns to spend latent computation there before emitting an answer.

Status

Not evaluated. No GSM8K accuracy was measured for this checkpoint, so there is no evidence here that it improves over the base model. Benchmark it before relying on it.

Usage

Loads as a plain Qwen2ForCausalLM — the run used thinking_use_mlp: false, so the checkpoint contains no modules beyond the base architecture.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "robinzixuan/LT-Tuning-Qwen2.5-1.5B-Instruct-GSM8K"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")

messages = [{"role": "user", "content": "Natalia sold clips to 48 friends in April, "
                                        "and then she sold half as many clips in May. "
                                        "How many clips did she sell altogether?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(inputs, max_new_tokens=512)[0], skip_special_tokens=True))

Training

Three stages of 1 epoch each, run back-to-back with the optimizer reset between stages:

Stage	Name	Mode	Thinking insertion prob
0	`stage0-cot`	`common`	0.0
1	`stage1-hidden-state`	`hidden_state`	0.85
2	`stage2-soft-fusion`	`soft_fusion`	0.95

Between stages the dataset is regenerated: every training example is scored with a forward pass (thinking_strategy: confidence) to pick where the <thinking> tokens go.


Hardware	8× H100 80GB
Precision	bf16, FlashAttention-2
Global batch	32 (4 per device × 8 GPUs × 1 grad accum)
Steps	702 (234 per stage)
LR	5e-5, warmup ratio 0.05, weight decay 0.01
Grad clipping	1.0
Wall clock	4h08m
Train data	GSM8K socratic, 7473 train / 1319 val

Loss

Final stage2-soft-fusion loss settled around 0.36 (grad_norm ~1.7, stable throughout).

The reported train_loss of 1.297 is the mean across all three stages and is not a meaningful convergence figure — the stages optimize different objectives, so their losses are not comparable to each other.

Limitations

Trained only on GSM8K grade-school word problems; nothing here transfers to other domains by default.
No held-out evaluation was run, on GSM8K or anything else.
Inherits the limitations and license terms of the Qwen2.5-1.5B-Instruct base model.