GoodStartLabs/qwen3-32b-gsl247-armF-v3-mixed-r4

Qwen3-32B · GSL-247 · arm F (v3-mixed-r4)

LoRA adapter trained on Lichess human games for the GSL-247 chess SFT ablation (form-vs-strategy axis). All four arms (A/B/C/F) share representation mix v3_mixed_history, N=200,000 records, identical hyperparameters; they vary only in the human-data filter applied upstream.

field	value
base model	Qwen/Qwen3-32B
PEFT	LoRA r=32, alpha=32, dropout=0.0, modules=all-linear
train slice	180,000 records (90/5/5 of 200k)
manifest hash	d4164be0b7b470a53e23fd52c9214d6f9b4aa1a6c8afde5e38f91a487947b8f4
manifest path	data/v3/armF_4x.manifest.json
representation	v3_mixed_history
epochs	1
optimizer	lr=0.001 (linear, warmup_ratio=0.03), batch=8
training method	sft
training infra	Together Fine-Tuning (job ft-4c44216b-f95f)
source commit	5b6e6b8

Arm definitions

• A: unfiltered human games
• B: top-decile by player Elo (per-time-control p90 floor)
• C: bottom-decile by player Elo (per-time-control p10 ceiling)
• F: arm A minus the top-decile (i.e., everyone except B)

The B-vs-C contrast is the pre-registered falsifier for whether better human data improves SFT on chess. F is an exploratory matched-N rebuild used to disambiguate A ≈ B ≈ C results from the v2 run.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-32B", device_map="auto")
tok  = AutoTokenizer.from_pretrained("Qwen/Qwen3-32B")
model = PeftModel.from_pretrained(base, "GoodStartLabs/qwen3-32b-gsl247-armF-v3-mixed-r4")

Eval contract

The pre-registered evaluation contract (rung-0 floor gates, rung-1 puzzle Elo ladder, B-vs-C falsifier thresholds) lives at eval/contract.yaml in the source repo. Every Inspect task stamps its contract_sha into the .eval metadata so any reported number is recoverable to the thresholds in force.