Qwen3-8B SDFT + math GRPO-LoRA (rollout 120)

Qwen3-8B (dense, 36 layers, 8.19B params). Fully merged bf16 weights — the LoRA adapter is folded into the base matrices, so this is a drop-in replacement for Qwen/Qwen3-8B-Base; no PEFT adapter loading required.

Training pipeline

1. Base: Qwen/Qwen3-8B-Base.
2. SFT: supervised fine-tuning on a balanced oracle mixture (math / retrieval-augmented QA / tool-use dialogue).
3. Math RL (this checkpoint): GRPO on math problems with a verifiable boxed-answer reward, training LoRA adapters (rank 128, alpha 128 ⇒ scaling 1.0) on q/k/v, o_proj, gate/up_proj and down_proj of every layer, for 120 rollouts. DAPO-style recipe (clip-higher 0.2/0.28, dynamic sampling, token-level policy-gradient loss) with truncated importance sampling for the train/rollout mismatch.

Trained with slime on Megatron-LM (TP=4, CP=1, bf16).

Honest evaluation note

Over these 120 rollouts the RL stage was stable but did not measurably improve the training-set pass@1 (0.558 → 0.578, linear slope +4e-5 per rollout against a per-rollout standard deviation of 0.027 — the change is inside the noise). Prompts solved 16/16 rose slightly (50.6 → 56.9 of 256 sampled) and prompts never solved were flat. Treat this as the SFT model plus a small, non-degrading RL delta, not as a demonstrated improvement over the SFT base. It is published for reproducibility and comparison, and it is the first checkpoint of this line whose frozen base weights are provably uncorrupted.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "willamazon1/Qwen3-8B-SDFT-Math-LoRA-pristine"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16).to("cuda")

prompt = "Question: What is 12*8?\nAnswer:"
ids = tok(prompt, return_tensors="pt").input_ids.cuda()
print(tok.decode(model.generate(ids, max_new_tokens=64, do_sample=False)[0][ids.shape[1]:],
                 skip_special_tokens=True))

This is a base-style (completion) model, not an instruction-tuned chat model: prompt it with Question: … \nAnswer: style completions rather than a chat template.

Limitations

Inherits the biases and knowledge cutoff of Qwen3-8B-Base. Math answers are not guaranteed correct — verify before relying on them.