Llama-ChemLink-Parser-8B-MTYS

ChemLink is a LoRA fine-tune of tokyotech-llm/Llama-3.1-Swallow-8B-Instruct-v0.3 for extracting chemical measurement values (MW, IC50, EC50, Yield) from scientific literature, with compound-name linkage for PubChem grounding and Graph RAG integration.

---

Background and Motivation

Target environment: CPU-only local hardware, no GPU required.

Chemical and pharmaceutical researchers frequently operate under security policies that prohibit cloud API usage. This model is designed to run on a standard CPU workstation (e.g., Core i7 / 24 GB RAM) via Ollama in GGUF format, suitable for overnight batch processing in network-restricted or air-gapped environments.

A critical requirement in this setting is compound-name linkage: downstream pipelines (PubChem grounding, Graph RAG) need to know not just the measurement value, but which chemical compound it belongs to. This requires the model to output a compound_name field alongside each extracted value.

Two prompt conditions were evaluated:

• Condition A (no instruction): prompt requests only type / value / unit; compound_name is not requested.
• Condition B (with instruction): prompt explicitly requests compound_name in addition to type / value / unit.

ChemLink outputs compound_name under both conditions. All comparison models output 0% compound_name without explicit instruction (Condition A).

---

Key Capability

ChemLink outputs compound_name alongside each extracted value under both prompt conditions, without dependence on explicit instruction.

{
  "chemical_entities": [
    {
      "compound_name": "linezolid",
      "measurements": [
        {"type": "Molecular Weight", "value": 337.35, "unit": "g/mol"}
      ]
    }
  ]
}

This stability reduces the risk of pipeline failures where a measurement value is extracted but cannot be linked to its source compound — a risk that depends on prompt design when using baseline models.

---

Model Overview

Item	Detail
Developer	MitzMitz / Ingenta AI
Base model	tokyotech-llm/Llama-3.1-Swallow-8B-Instruct-v0.3
Training tool	unsloth + TRL (SFTTrainer)
Quantization	4-bit NF4 (QLoRA)
LoRA config	r=16, alpha=32, dropout=0, bias=none
Max seq length	2048
Local deployment	Ollama (GGUF q5_K_M) — CPU only, no GPU required
Supported languages	Japanese, English
License	Llama 3.1 Community License

---

Usage

Local CPU Inference (Ollama — Primary Use Case)

# モデル登録
ollama create llama-chemlink-parser-8b-mtys -f Modelfile

# 推論
ollama run llama-chemlink-parser-8b-mtys

Modelfile example (Llama 3.1 chat template):

FROM ./Llama-3.1-Swallow-8B-Instruct-v0.3.Q5_K_M.gguf

TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>

{{ .System }}<|eot_id|>{{ end }}<|start_header_id|>user<|end_header_id|>

{{ .Prompt }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{{ .Response }}<|eot_id|>"""

PARAMETER temperature 0
PARAMETER num_ctx 2048
PARAMETER num_predict 256

Inference (Colab / GPU)

import torch, json, re
from unsloth import FastLanguageModel
from google.colab import userdata

HF_TOKEN = userdata.get('HF_TOKEN')

SYSTEM_PROMPT = (
    "You are a chemical data extraction assistant. "
    "Extract measurements from the given text and return a JSON object. "
    "The object must have a 'chemical_entities' array. "
    "Each element must have: compound_name (string), "
    "measurements (array of objects with type/value/unit). "
    "If no target measurement is found, return {\"chemical_entities\": []}. "
    "Output only the JSON object, no explanation."
)

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name     = "MitzMitz/Llama-ChemLink-Parser-8B-MTYS",
    max_seq_length = 2048,
    dtype          = None,
    load_in_4bit   = True,
    token          = HF_TOKEN,
)
FastLanguageModel.for_inference(model)

def extract(text):
    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user",   "content": text},
    ]
    input_ids = tokenizer.apply_chat_template(
        messages, tokenize=True,
        add_generation_prompt=True, return_tensors="pt"
    ).to("cuda")
    with torch.no_grad():
        output = model.generate(
            input_ids, max_new_tokens=256,
            temperature=0.0, do_sample=False,
            pad_token_id=tokenizer.eos_token_id,
        )
    return tokenizer.decode(
        output[0][input_ids.shape[1]:], skip_special_tokens=True
    ).strip()

print(extract("The compound linezolid has a molecular weight of 337.35 g/mol."))

---

Training Configuration

Parameter	Value
per_device_train_batch_size	1
gradient_accumulation_steps	16
num_train_epochs	2
learning_rate	2e-4
warmup_steps	10
lr_scheduler_type	cosine
fp16 / bf16	auto-detected
optimizer	adamw_8bit (unsloth default)
save_strategy	steps (save_steps=20)

---

Training Data

File	Total	MW	IC50	EC50	Yield	Negative	Source
phase6_train_mix	3,763	2,283	717	0	44	719	PubChem / ChEMBL / ORD
additional_ec50_yield	2,534	0	0	1,000	1,000	534	ChEMBL / ORD
additional_yield_table	621	0	0	0	500	121	ORD
additional_mw_unit_fix	120	84	16	0	0	20	PubChem
additional_phase5	740	17	115	22	425	161	ChEMBL / ORD / PubChem
Total	7,778	2,384	848	1,022	1,969	1,555

Negative samples (1,555 records, 20.0%) contain [] as output, representing texts where no target measurement exists.

Data licenses:

• ORD: CC-BY-SA 4.0
• ChEMBL: CC-BY-SA 3.0 (EMBL-EBI)
• PubChem: Public Domain (NCBI/NIH)

---

Evaluation

Evaluation Dataset

Indicator	n (chemfmt evaluation)	Source
MW	125 per model	PubChem (PMID-verified)
Yield	125 per model	ORD (PMID-verified)
IC50	125 per model	ChEMBL → PubMed Abstract
EC50	125 per model	ChEMBL → PubMed Abstract

Evaluation used true_eval_all_pmid_clean.jsonl (2,963 records total; PMID-verified, zero training data contamination). 500 samples per model per condition (125 per indicator, stratified, RANDOM_SEED=42).

Local CPU / Ollama Evaluation (q5_K_M GGUF)

Prompt format: chemical_entities structured output. Parameters: temperature=0.0, num_predict=256, num_ctx=2048.

MW extraction — Condition A (no instruction):

Model	MW parsed (n/125)	compound_name	PubChem resolved	MW accuracy*
ChemLink q5_K_M	124/125	100%	60.5%	100%
Swallow-base q5_K_M	124/125	100%†	60.5%	100%
Mistral-7B q5_K_M	123/125	0%	—	100%

MW extraction — Condition B (with compound_name instruction):

Model	MW parsed (n/125)	compound_name	PubChem resolved	MW accuracy*
ChemLink q5_K_M	125/125	100%	64.8%	100%
Swallow-base q5_K_M	125/125	100%	64.0%	100%
Mistral-7B q5_K_M	67/125 ⚠️	100%‡	47.8%	100%

* MW accuracy = fraction of extracted values within ±1% of gold-standard truth value.
† Swallow-base compound_name output under Condition A is an artifact of the Ollama chat template (see Limitations); this behaviour is not observed in the Colab/GPU evaluation of the same base model.
⚠️ Mistral-7B Condition B: low n (67/125) due to chemical_entities format incompatibility causing truncation or empty outputs (see Limitations).

Colab GPU / NF4 Reference Results

Parameters: temperature=0.0, max_new_tokens=256.

MW extraction:

Model	Condition	MW parsed (n/125)	compound_name	PubChem resolved
ChemLink NF4	no instruction	123/125	100%	59.3%
ChemLink NF4	with instruction	120/125	100%	63.3%
Swallow-base	no instruction	123/125	0%	—
Swallow-base	with instruction	124/125	100%	64.5%
Mistral-7B	no instruction	112/125	0%	—
Mistral-7B	with instruction	32/125 ⚠️	100%‡	68.8%‡

‡ When values were successfully extracted.

PubChem Grounding Summary

Of all compound names output by ChemLink (Condition A, local CPU, MW):

• PubChem resolution rate: ~60–65%
• Non-resolved names are typically IUPAC systematic names, synthesis intermediates, or proprietary codes not registered in PubChem.

---

Limitations

compound_name in real abstracts: Real PubMed abstracts often use codes ("compound 3", "2b") rather than IUPAC names. The model outputs whatever name appears in the source text; PubChem grounding success depends on name quality.

Mistral-7B format incompatibility: Mistral-7B-Instruct-v0.2 produced valid MW extractions in only 25–51% of cases under the chemical_entities prompt format, compared to >95% for ChemLink and Swallow-base. This is attributed to the [INST] chat template being incompatible with longer structured JSON output, causing truncation. Mistral-7B is not recommended for chemical_entities format inference.

Swallow-base Ollama Condition A artifact: Swallow-base q5_K_M showed unexpected compound_name output under Condition A (no instruction) via Ollama, which was not observed in the same model evaluated via Colab/GPU. This is attributed to a difference in chat template handling between Ollama and apply_chat_template. The Colab result (0% compound_name) is considered more reliable.

IC50 / EC50: Across all models evaluated, IC50/EC50 correct-extraction rates were below 2%. This reflects a limitation of the unified structured-output evaluation protocol, not a model-specific capability gap.

Inference environment differences: Colab GPU: temperature=0.0, max_new_tokens=256, apply_chat_template. Local Ollama: temperature=0.0, num_predict=256, Modelfile TEMPLATE. Cross-environment comparisons should account for these differences.

---

Intended Use

• Automated extraction of MW / Yield from chemical literature in network-restricted, CPU-only local environments
• Compound-name to measurement-value association for PubChem grounding and Graph RAG pipelines
• Overnight batch processing on CPU-only hardware without cloud API dependency

Out-of-Scope Use

• Medical diagnosis or legal judgment
• Domains outside chemistry and chemical biology
• IC50/EC50 extraction (see Limitations)

---

Base Model Reference

Model	License
tokyotech-llm/Llama-3.1-Swallow-8B-Instruct-v0.3	Llama 3.1 Community License
meta-llama/Llama-3.1-8B-Instruct	Llama 3.1 Community License

---

License

Licensed under the Llama 3.1 Community License. Copyright (C) Meta Platforms, Inc. All Rights Reserved.

---

Framework Versions

Library	Version
unsloth	2026.5.2
PEFT	0.19.1
Transformers	5.5.0
PyTorch	2.10.0
TRL	0.24.0
Datasets	4.3.0