Downloads · 30 days
0
leniencybench-qwen3b-outputs
shreyas-garg/leniencybench-qwen3b-outputs
leniencybench-qwen3b-outputs is a machine learning model from shreyas-garg. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as mit.
LoRA adapter for Qwen 2.5 3B-Instruct, fine-tuned on LeniencyBench, an OpenEnv-compliant environment that measures and trains out leniency bias in LLMs — the systematic failure to apply mid-context rule tightenings.
Access
Public
Updated Apr 26, 2026
Repo size
143 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors240 MB · 91%
At a glance
- Library
- peft
- License
- mit
- Access
- Public
- Created
- Apr 25, 2026
- Updated
- Apr 26, 2026
- SHA
- a73e4024
Base models
Spaces
Model card
From the Hugging Face model README
LeniencyBench — Qwen 2.5 3B SFT-trained adapter
LoRA adapter for Qwen 2.5 3B-Instruct, fine-tuned on LeniencyBench, an OpenEnv-compliant environment that measures and trains out leniency bias in LLMs — the systematic failure to apply mid-context rule tightenings.
Headline result
| Stage | Tightening accuracy | Loosening accuracy | Drift-sensitive overall |
|---|---|---|---|
| Pre-training (Qwen 2.5 3B base) | 0.0 % (0/23) | 21.4 % (3/14) | 11.8 % (2/17) |
| Post-SFT (this adapter) | 91.3 % (21/23) | 71.4 % (10/14) | 88.2 % (15/17) |
200 held-out samples on episode seeds 10000–10039 (disjoint from the 0–799 training seeds).
The result reproduces exactly across two independent training runs (a10g + a100), giving us confidence the gap closure is the env's signal, not run-to-run noise.
What's in this repo
| File | What it is |
|---|---|
lora_adapters/ | Final LoRA adapter (post-SFT — this is what you load) |
lora_adapters_sft/ | Same adapter, kept under both names for clarity |
evals.json | Pre / post-SFT eval snapshots with per-direction breakdown |
sft_log.json | Full SFT training log (loss per step, ~1000 steps) |
v6_full_logs.txt | Raw stdout from a parallel run that confirmed the same result |
baseline_direction_split.png | Cross-model baseline plot |
Loading the adapter
from unsloth import FastLanguageModel
from peft import PeftModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="unsloth/Qwen2.5-3B-Instruct",
max_seq_length=4096,
load_in_4bit=True,
)
model = PeftModel.from_pretrained(
model, "shreyas-garg/leniencybench-qwen3b-outputs",
subfolder="lora_adapters",
)
FastLanguageModel.for_inference(model)
Training recipe
- Base: unsloth/Qwen2.5-3B-Instruct (4-bit)
- LoRA: rank 16, alpha 16, dropout 0, all 7 linear projections targeted
- Optimizer: SFTTrainer (HF TRL 0.24), lr 2e-4, batch 4, grad accum 4, bf16
- Data: 16,000 per-step samples drawn from 800 LeniencyBench episodes (seeds 0–799)
- Compute: 1 epoch in ~100 min on A100-SXM4-80GB
- Held-out eval: seeds 10000–10039, capped at 200 samples
Read more
- Live env (run it yourself): shreyas-garg/drift-env
- Source code (HF mirror): shreyas-garg/leniencybench
- Source code (GitHub): shreyas-garg/OpenEnv
Built for the Meta PyTorch × Hugging Face OpenEnv Hackathon (Round 2).
- Library
- peft
- License
- mit
- Languages
- en
- Created
- Apr 25, 2026
- Updated
- Apr 26, 2026