RockToken

qwen3_30b_a3b_to_4b_onpolicy_code_5k

README

License: apache-2.0

Models

Table with columns: Role, Model
Role	Model
Student	RockToken/qwen3_30b_a3b_to_4b_offpolicy_20k
Teacher	Qwen/Qwen3-30B-A3B-Instruct-2507 (MoE)

enable_thinking=False throughout.

Training data

Source: open-thoughts/OpenThoughts3-1.2M, domain == "code" slice
5,000 single-user-turn prompts; 70 exceeding prompt_max_len=2560 were filtered, leaving 4,930
Only the prompts are used; on-policy KD never reads the dataset's reference answers

Training setup

Framework: KDFlow — FSDP2 + SGLang rollout, Ray-orchestrated GPU co-location with sleep/wakeup. Same pipeline revision as the math-domain onpolicy_5k_src* runs.

Hardware: 1× node, 4× H100 (94 GB), 33 h 09 min wall-clock (~133 GPU-hours).

Key hyperparameters

Table with columns: Group, Value
Group	Value
Backend	fsdp2, bf16, gradient ckpt on
Epochs	1 (2,465 rollout iterations)
Train batch	4 (micro 1)
Learning rate	2e-6, cosine, warmup 5%
KD ratio	1.0
KD loss	reverse KL (`rkl`)
KD algorithm	`vanilla_kd`
Rollout engine

prompt_max_len is 2560 instead of the math runs' 800: code prompts carry problem statements and starter code (p99 ≈ 2,356 tokens).

Note on the LR schedule: as in the earlier runs of this series, the scheduler's horizon is sized to 4× the actual number of updates, so the cosine decay is only partially traversed — the LR moves from 2.00e-6 to 1.79e-6 over the run rather than reaching zero. Kept as is for comparability within the series.

Training dynamics

Means over 500-step blocks (per-step values are noisy at batch 4):

Table with columns: steps, loss (reverse KL), mean gen length
steps	loss (reverse KL)	mean gen length
1–500	0.190	6,814
501–1000	0.178	6,611
1001–1500	0.175	6,471
1501–2000	0.166	6,647
2001–2465	0.166	6,378

The loss starts near 0.19 rather than the ~3 seen when distilling from a raw instruct model — the student has already been through off-policy KD against the same teacher, so on-policy training starts close and grinds out the remaining distribution gap.

⚠ Truncation caveat

Code responses run long, and a substantial share of rollouts hit the 8,000-token generation cap: per-step mean lengths sit at p50 = 7,069 against the cap, and in 20% of steps the whole batch was truncated (step-mean ≥ 7,990). Truncated rollouts contribute KD signal only for their first 8,000 tokens, and the student receives no supervision on how the teacher would have ended those responses. Treat comparisons against the science and math checkpoints (which trained essentially truncation-free) with this in mind.

Intended use

Research on distillation dynamics, in particular whether on-policy KD gains transfer across domains (math-trained off-policy base → code on-policy round). Domain: code (OpenThoughts-3 code split).

Limitations

The on-policy round is code-only; not tuned for chat or safety.
enable_thinking=False — this student does not emit thinking traces.
Heavy generation-length truncation during training; see the caveat above.
Single checkpoint at the end of one epoch; no intermediate checkpoints were kept.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

RockToken

Model Tree

Base

RockToken/qwen3_30b_a3b_to_4b_offpolicy_20k

Fine-tuned

this model

Input Modalities

Text

Output Modalities