Qwen3-8B, multi-stage SFT (SDFT) checkpoint with a tau-bench RL LoRA adapter merged in. Full weights, ready to load with transformers / SGLang / vLLM — no PEFT needed.

This is iteration 240 of the same run that produced willamazon1/sdft-tau-lora-iter160 (80 more RL steps).

Lineage

Stage	What
Base	`Qwen/Qwen3-8B-Base`
SDFT	multi-stage SFT chain: Math → Sea → Search → TauSFT → Tau-IF
RL	GSPO on tau-bench `retail` (train split), LoRA-only (base frozen), iteration 240

RL / LoRA configuration

adapter: rank 16, alpha 32 (scaling = alpha/r = 2.0), dropout 0.0
targets: linear_qkv, linear_proj, linear_fc1, linear_fc2 on all 36 layers (144 modules, 288 tensors)
advantage estimator GSPO, KL loss 0.01 (low_var_kl), eps-clip 0.2/0.25
Adam, lr 1e-6 constant, weight decay 0.01
TIS off (its sequence-level rejection veto zeroes essentially every LoRA sequence)
user simulator: GLM-4.7-Flash served in-cluster (matches tau-bench's LLM-user setup)

How it was exported

The adapter was merged in Megatron parameter space (W += 2.0 · B·A per LoRA'd module, into the frozen base weights carried by the same iter_0000240 torch_dist checkpoint), then converted to HuggingFace safetensors. Merging before conversion avoids having to un-fuse the GQA-interleaved QKV and the gate/up split by hand. LayerNorm weights are untouched — the LoRA delta applies to the post-LN matmuls only.

Verification performed on this export:

adapter health gate: 144/144 modules have non-zero lora_B (max |lora_B| = 1.383e-04), so the adapter is genuinely trained rather than sitting at its zero init
relative delta size ‖ΔW‖/‖W‖: min 4.80e-05, median 6.35e-05, max 1.40e-04 (larger than iter 160's 3.76e-05 / 5.01e-05 / 1.11e-04, as expected after 80 more steps)
0 non-finite tensors across all 4 shards; 399 tensors, 15.26 GiB, index complete

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "willamazon1/sdft-tau-lora-iter240"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")

Note this is derived from a base (non-instruct) Qwen3 checkpoint plus SFT/RL stages; use the same prompt format as the tau-bench agent it was trained with rather than assuming a generic chat template.