mars-v-ft-checkpoints

Fine-tuning checkpoints — MARS V

LoRA adapters saved during training, for trajectory / PCA analysis. Task: translating a natural-language story into a rigid formal grammar, graded mechanically.

All runs: LoRA r=16, alpha=16, dropout 0, on q,k,v,o,gate,up,down_proj in every layer. lr 2e-4 cosine, 3% warmup, seed 0, 3 epochs. Every run includes step-0 (initialized, untrained), so each trajectory has its own origin.

The 8B and the 32B each have both recipes — same model, same LoRA config, only the training data differs — and the outcomes are opposite: grammar-only training perfects syntax and destroys accuracy (32B 34% → 3%), task pairs teach the task (32B 34% → 99.7%).

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
model = PeftModel.from_pretrained(base, "llama-3.1-8b_task-pairs/step-500")

Weight deltas

dW = (alpha/r) * B @ A, rank-16 by construction — the factors are the compact form (168 MB factored vs ~2 GB dense per 8B checkpoint).

from safetensors import safe_open

with safe_open("step-500/adapter_model.safetensors", framework="pt") as sf:
    A = sf.get_tensor("base_model.model.model.layers.0.self_attn.q_proj.lora_A.weight")
    B = sf.get_tensor("base_model.model.model.layers.0.self_attn.q_proj.lora_B.weight")
dW = B.float() @ A.float()

Each folder also has trajectory.npz — steps, modules, norms (‖dW‖ per module per checkpoint) — enough to see where and when a model changed without downloading any adapter.

Notes

• qwen3-32b_task-pairs stops at step-900, not a full 3 epochs: its training container was lost to a network failure at ~2.6 epochs. Its accuracy curve is flat from step 500, so little was left to change.
• Adapters only, no base weights redistributed. The two Llama runs inherit the Llama 3.1 Community License; the Qwen3 and Ministral runs are Apache 2.0.