camgeodesic/olmo3-7b-instruct-only

OLMo-3 7B Instruct-Only (GRPO)

Fine-tuned from allenai/OLMo-3-1025-7B using GRPO (Group Relative Policy Optimization) on instruction-following tasks.

Training Details

• Base model: allenai/OLMo-3-1025-7B
• Training method: GRPO with RL-Zero (no supervised warmup)
• Dataset: allenai/Dolci-RLZero-IF-7B (IFEval instruction-following)
• Config: if_valley_thinker — valley length penalty (512–4096 token sweet spot) + think token reward shaping
• Chat template: OLMo thinker (prefills <think> tag for chain-of-thought reasoning)
• Precision: bfloat16
• Hardware: 2 nodes × 4 NVIDIA GH200 120GB GPUs (Isambard-AI)

Reward Components

Component	Description
IFEval verifiable reward	Binary per-constraint score for instruction-following
Valley length penalty	Penalizes responses <512 or >4096 tokens (coeff: -0.001)
Think tag reward	+0.125 for correct </think> closure
Think length penalty	-0.1 if thinking block <10 words

Performance (late-stage averages)

Metric	Value
IFEval correct rate	0.88
Training reward	6.36
Think word count	~886 words
Sequence length	~1353 tokens

Checkpoints

Each training checkpoint is available as a separate branch/revision:

• main — step 3800 (latest)
• step_600 through step_3600 — intermediate checkpoints (every 200 steps)

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load latest
model = AutoModelForCausalLM.from_pretrained("camgeodesic/olmo3-7b-instruct-only")

# Load specific checkpoint
model = AutoModelForCausalLM.from_pretrained("camgeodesic/olmo3-7b-instruct-only", revision="step_2000")