haielab/DeepSeek-Prover-V2-7B-conjecture-base-FineTune-20250724_0955

DeepSeek‑Prover‑V2‑7B · LoRA Adapter

This repository hosts a LoRA adapter fine‑tuned on top of deepseek-ai/DeepSeek-Prover-V2-7B using 🤗 trl’s SFTTrainer.

---

Training Setup

Hyper‑parameter	Value
Learning rate	2 × 10⁻⁴
Batch size / device	16
Gradient accumulation steps	1
Effective batch size	16
Epochs	1
Scheduler	linear
Warm‑up ratio	0.03
Weight decay	0.01
Seed	42
Sequence length	1 792
Flash‑Attention‑2	✅ (use_flash_attention_2=True)

LoRA configuration

Setting	Value
Rank r	16
α	32
Dropout	0.05
Target modules	all linear layers
Modules saved	embed_tokens, lm_head
Bias	none

RoPE scaling: YARN, factor = 16.0, β_fast = 32.0, β_slow = 1.0

Training was performed on GPUs with bfloat16 precision (torch_dtype=torch.bfloat16).

---

Loss Curve

---

Usage

from transformers import AutoTokenizer, AutoPeftModelForCausalLM

model = AutoPeftModelForCausalLM.from_pretrained(
    "your‑username/DeepSeek-Prover-V2-7B-conjecture-chat-new-config-20250724_0955",
    trust_remote_code=True,
)
tok = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-Prover-V2-7B", trust_remote_code=True)

prompt = "Prove that the sum of two even numbers is even."
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=128)
print(tok.decode(out[0], skip_special_tokens=True))