A 4B math-distilled model. Student fine-tuned from RockToken/qwen3_30b_a3b_to_4b_onpolicy_5k_src20k-25k_freeze_rock via on-policy reverse-KL distillation against a Qwen3-30B-A3B teacher, using the token_freeze_kd algorithm to mask a 98-token "freeze list" out of the KD loss.

This is the treatment arm ("rock" token set) of an A/B experiment on the effect of selectively freezing certain tokens during on-policy KD.

Training


Student (init)	`RockToken/qwen3_30b_a3b_to_4b_onpolicy_5k_src20k-25k_freeze_rock`
Teacher	`Qwen/Qwen3-30B-A3B-Instruct-2507`
Data	OpenThoughts3 math prompts — 10k cumulative (5k src20k-25k → continued 5k src25k-30k)
Source file	`openthoughts_prompt_math_5k_src25k-30k.jsonl (continuation)`
KD algorithm	`token_freeze_kd` (see KDFlow)
KD loss	reverse-KL, kd_temperature=1.0, kd_ratio=1.0
Freeze list	98 token IDs from `rock.json` (treatment set)
Freeze weight	0.0 (loss on these tokens is fully zeroed)
Backend	FSDP2, bf16, gradient checkpointing
Topology	4× H100 (1 node), teacher TP=4, rollout TP=2
Optimizer	AdamW, lr=2e-6 (5% warmup, cosine→1e-8), max_norm=1.0
Batch	train_batch_size=4, micro=1, n_samples_per_prompt=4
Steps	1 epoch, 2500 steps
Rollout	response cap 1024 tokens, temperature=1.0, top_p=1.0
Chat template	applied; `enable_thinking=False` (Instruct-2507 is non-thinking)

A/B siblings

5k variants (this slice only):
- RockToken/qwen3_30b_a3b_to_4b_onpolicy_5k_src20k-25k_freeze_random — control
- RockToken/qwen3_30b_a3b_to_4b_onpolicy_5k_src20k-25k_freeze_rock — treatment
10k variants (5k + continued 5k on src25k-30k):
- RockToken/qwen3_30b_a3b_to_4b_onpolicy_10k_src20k-30k_freeze_random — control
- RockToken/qwen3_30b_a3b_to_4b_onpolicy_10k_src20k-30k_freeze_rock — treatment

Use

python

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("RockToken/qwen3_30b_a3b_to_4b_onpolicy_10k_src20k-30k_freeze_rock")
model = AutoModelForCausalLM.from_pretrained("RockToken/qwen3_30b_a3b_to_4b_onpolicy_10k_src20k-30k_freeze_rock", torch_dtype="bfloat16")