ssfc/pcf-ministral3-base-detail-to-detail-lora

PCF Ministral-3-14B Base Detail-to-Detail LoRA

This repository contains a LoRA adapter for mistralai/Ministral-3-14B-Base-2512 trained for the Past-Creates-Future (PCF) algorithm-idea generation workflow.

The model is trained on the same detail-to-detail PCF task as the Qwen3 detail-to-detail adapter:

existing method detail + method delta -> new method detail

It is intended as an alternative base-model experiment for structured algorithm-idea generation.

Model Details

• —Base model: mistralai/Ministral-3-14B-Base-2512
• —Adapter type: LoRA / PEFT
• —Task type: causal language modeling
• —LoRA rank: 16
• —LoRA alpha: 16
• —LoRA dropout: 0.05
• —Target modules: all major linear projection modules (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj)
• —Training dtype: bf16

Intended Use

This adapter is designed for algorithm-research ideation. It can be used to:

• —generate structured descriptions of candidate algorithm variants;
• —apply mechanism-level method deltas to existing algorithms;
• —compare PCF behavior across different 14B-scale base models;
• —produce implementation-oriented method sketches for later human validation.

The model is not a replacement for experimental evaluation. It proposes candidate mechanisms; a researcher still needs to implement and benchmark them.

Out-of-Scope Use

Do not use this adapter as evidence that a generated algorithm is correct, optimal, safe, or empirically better. It may generate plausible details that are incomplete or wrong.

Input Format

The training format uses structured method details with these fields:

problem_definition
method_name
method_type
core_idea
key_components
input_output
optimization_objective
algorithm_flow
assumptions
limitations
novelty_claim
prior_work_relation
transferable_delta_hints

The method delta contains fields such as:

new_title
problem_shift
method_delta
component_changed
mechanism_added
mechanism_removed
abstraction_shift
why_improves
reuse_pattern
delta_keywords

The expected output is a valid JSON object with the method-detail fields.

Usage

Ministral-3 models may require recent transformers support. The local PCF inference script uses AutoModelForImageTextToText when the model config reports model_type == "mistral3".

import torch
from peft import PeftModel
from transformers import AutoConfig, AutoModelForCausalLM, AutoModelForImageTextToText, AutoTokenizer, BitsAndBytesConfig

base_model = "mistralai/Ministral-3-14B-Base-2512"
adapter = "ssfc/pcf-ministral3-base-detail-to-detail-lora"

config = AutoConfig.from_pretrained(base_model, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(
    base_model,
    trust_remote_code=True,
    fix_mistral_regex=True,
)
model_cls = AutoModelForImageTextToText if getattr(config, "model_type", None) == "mistral3" else AutoModelForCausalLM
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
)
model = model_cls.from_pretrained(
    base_model,
    device_map="auto",
    quantization_config=quantization_config,
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

For the full prompt construction used by this project, see finetune/inference_detail.py in the Past-Creates-Future repository.

Training Data

Training data file:

deepseek_method_delta_details_v4_flash_30k5562_detail_to_detail_conf0p6.jsonl

The dataset was built from algorithm papers. The pipeline extracts structured method details for old and new methods, summarizes method deltas, filters examples by confidence, and formats examples as:

existing detail + delta detail -> new detail

This version uses a 30k-character paper extraction context and a confidence threshold of 0.6.

Training Configuration

• —Epochs: 3
• —Per-device batch size: 1
• —Gradient accumulation steps: 16
• —Learning rate: 1e-4
• —Scheduler: cosine
• —Max sequence length: 4096
• —Optimizer: paged AdamW 8-bit
• —Gradient checkpointing: enabled
• —Liger kernel: enabled
• —Hardware used locally: single NVIDIA GPU with 24 GB VRAM

Evaluation

This adapter was compared against the Qwen3 detail-to-detail adapter on held-out PCF examples. It is useful as a base-model comparison point, while the Qwen3 detail-to-detail adapter has been the stronger default choice in local PCF experiments so far.

The uploaded adapter was downloaded back from Hugging Face and verified against the local upload source by SHA256 for adapter_config.json, adapter_model.safetensors, and README.md.

Limitations

• —This is an experimental adapter for structured algorithm ideation.
• —It can generate plausible but unverified algorithm mechanisms.
• —It may require newer transformers support than the Qwen3 adapters.
• —Output should be reviewed and experimentally validated before being used in research claims.

Framework Versions

• —PEFT 0.20.0