evenfarther/DeepSeek-R1-Distill-Qwen-14b-chemical-synthesis-adapter-high-TPR

DeepSeek-R1-Distill-Qwen-14B Chemical Synthesis Classifier — LoRA Adapter (Epoch 2)

Model Overview

This repository contains the LoRA adapter for a DeepSeek‑R1‑Distill‑Qwen‑14B model fine‑tuned to classify chemical synthesizability (P = synthesizable, U = unsynthesizable). Training uses QLoRA on an imbalanced P/U dataset; evaluation scores each example with logsumexp(P) − logsumexp(U) at the final token under the SFT‑aligned prompt.

• Task: Binary classification (P vs U)
• Objective: QLoRA with class‑imbalance handling
• Max sequence length (eval): 180 tokens
• Dataset (train): train_pu_hem.jsonl (316,442 samples)
• Dataset (val): validate_pu_hem.jsonl (79,114 samples)
• Adapter size: ~1.10 GB (adapter_model.safetensors)
• Checkpoint: Epoch 2 — checkpoint-496 (best TPR among evaluated epochs)

The checkpoint includes a chat_template.jinja to ensure prompt formatting matches SFT conditions.

Prompt & Assistant Prefill

Evaluation and recommended usage preserve the SFT formatting and avoid think tokens:

• System: You are a helpful assistant for P/U classification of synthesizability.
• User: Dataset-provided query. In the training corpus it follows the pattern: As chief synthesis scientist, judge {composition} critically. Avoid bias. P=synthesizable, U=not:
• Assistant (prefill): empty assistant start; do not insert <think>

Implementation notes for faithful scoring/inference:

• Build inputs via the chat template; drop the final assistant label from dataset inputs and tokenize with an empty assistant turn.
• Use add_generation_prompt=False and read logits right after the assistant start (trim the trailing EOS if present).
• Force attn_implementation="eager" for stability.

Validation Metrics (Epoch 2 — Best)

Metric	Value
TPR (P Recall)	0.9901
TNR (U Specificity)	0.8160

Source: sft/deepseek_14b_sft_outputs/validation_metrics.csv (loader: transformers).

How to Load (Transformers + PEFT)

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

base = "unsloth/DeepSeek-R1-Distill-Qwen-14B-bnb-4bit"  # Base model repo or equivalent local path
adapter = "<this-repo-or-local-checkpoint-dir>"        # This adapter (checkpoint-496)

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
)

tok = AutoTokenizer.from_pretrained(base, use_fast=True)
tok.padding_side = "right"
if tok.pad_token is None:
    tok.pad_token = tok.eos_token

model = AutoModelForCausalLM.from_pretrained(
    base,
    quantization_config=bnb,
    device_map="auto",
    attn_implementation="eager",
)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

# Build messages using the included chat template
messages = [
    {"role": "system", "content": "You are a helpful assistant for P/U classification of synthesizability."},
    {"role": "user", "content": "As chief synthesis scientist, judge FeCrNiMn high-entropy alloy critically. Avoid bias. P=synthesizable, U=not:"},
    {"role": "assistant", "content": None},  # empty assistant prefill (no <think>)
]

ids = tok.apply_chat_template(messages, tokenize=True, add_generation_prompt=False, return_tensors="pt").to(model.device)
out = model(input_ids=ids, use_cache=False)
logits = out.logits[0, -1]
# Compare logsumexp over token ids for P vs U to obtain the score.

Training Data

• Train split: train_pu_hem.jsonl — 316,442 samples
  • Label distribution: 34,685 P (10.96%), 281,757 U (89.04%)
• Validation split: validate_pu_hem.jsonl — 79,114 samples
  • Label distribution: 8,674 P (10.96%), 70,440 U (89.04%)
• Format: JSONL with messages = [{role: "user", content: pattern above}, {role: "assistant", content: "P"|"U"}]
• Training injects the fixed system message shown above and tokenizes with add_generation_prompt=False, appending EOS.

Training Setup (Summary)

• Base model: Unsloth "DeepSeek‑R1‑Distill‑Qwen‑14B‑bnb‑4bit" (4‑bit NF4)
• Fine‑tuning: QLoRA via PEFT
  • Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
  • LoRA config: r=32, alpha=32, dropout=0.0
• Objective: P/U‑only Focal Loss applied to the last P or U token
  • gamma=2.0, alpha_P=8.12, alpha_U=1.0
• Tokenizer: right‑padding; pad_token = eos_token if undefined
• Saving: periodic checkpoints; checkpoint-496 corresponds to Epoch 2

Dataset Sources

The training/validation corpus combines multiple public sources and internal curation:

• P/U labelled data from J. Am. Chem. Soc. 2024, 146, 29, 19654-19659 (doi:10.1021/jacs.4c05840)
• High-entropy materials data from Data in Brief 2018, 21, 2664-2678 (doi:10.1016/j.dib.2018.11.111)
• Additional candidates via literature queries and manual screening of high-entropy materials

After de-duplication across sources, the validation split contains 79,114 samples with an imbalanced label ratio (~11% P / 89% U).

VRAM & System Requirements

• GPU VRAM: ≥16 GB recommended (4‑bit base + adapter)
• RAM: ≥16 GB recommended for tokenization and batching
• Libraries: transformers, peft, bitsandbytes (evaluation uses transformers loader)
• Set attn_implementation="eager" to avoid SDPA instability

Limitations & Notes

• The adapter targets chemical synthesizability judgments; generalization outside this domain is not guaranteed.
• For consistent results, use the included chat_template.jinja and avoid inserting <think> tokens.
• Do not mutate the base model config.json (e.g., model_type), which can reinitialize weights and corrupt metrics.